Exhibit four · the licence ledger · living

The licence ledger.

exhibit four Workshop notes
Published 2026-08-11
rows last read 2026-10-22 · page updated 2026-10-24 (UTC)
a small (human) team and a fleet of AI agents

Every model our products run — and every model we benched on the way — with its licence read first-hand and stamped with the date we last read it. Not a summary of someone's summary: the licence file that ships with the weights, the platform's raw permission record, the vendor's current text. When a clause moves, the row moves, and the change is named here.

ask about this page → assistant.strata2signal.com · in beta, still being tested

hardware on this page: RTX PRO 6000 Blackwell (96 GB) → research.strata2signal.com/hardware/ · the roster is in beta, still being tested

the short version

Every model we run or bench, its licence read from the actual text — the page the others cite instead of restating.

12,442 words, about 57 minutes to read.

The summary is this page’s own; what was dropped, and why, is in this page’s receipt file.

How to read this page

Each row carries six columns — the model, the seat it holds here, the licence, the day we last read it, what that licence asks of us, and what it permits you. The method for each read is named once per table, above it, so you can repeat the read yourself. The date on each row is the day that licence text was last read in full by us — not the day the row was written. First things first, the fact every row stands on: our products ship zero model weights. In the shipped products, every model is one you choose and download yourself, on your machine, under that model's own licence. And where a licence withholds hosted-service use, we treat that as binding on anyone operating a server, ourselves included — that is why one painter below — Juggernaut-XL v9 — is benched and never wired.

We are not lawyers, and this is research notes, not legal advice — each licence is its own authority, linked or named so you can read it yourself. Corrections welcome: hello@strata2signal.com. Model, company, and platform names on this page are the trademarks of their respective owners, used only to identify what we read — no affiliation or endorsement is implied, and no licence here grants us any trademark right.

The brains · chat, vision, memory · all read 2026-08-11

Read via ollama show <tag> --license — the licence file packaged with the weights themselves, read in full, not skimmed for a name.

gemma4:12b (q4)the narrator — the locked co-resident chat seatread 2026-08-11
licenceApache 2.0
what it asks of us§4 notice duties if redistributing
what it permits yourun it anywhere, commercially, royalty-free — no use policy, no output claims, and no trademark rights come with it (§6)
gemma4:26b (a4b)the 96G workstation posture seat (the locked co-resident default remains 12b)read 2026-08-11
licenceApache 2.0
what it asks of ussame §4 duties
what it permits yousame — trademarks withheld, patent grant self-terminates on suit; no use policy, no output claims
qwen3-vl:2bno seat today — this row named it the world's eyes, the game world's vision intake, on 2026-08-11; that seat has been empty since 2026-09-08, and the last model in it, MiniCPM-V 4.5, has its own row in the new brains section belowread 2026-08-11
licenceApache 2.0
what it asks of ussame §4 duties
what it permits yousame — and no licence-level limit on vision use; our refuse-identifiable-people rule is our own architecture test, not a licence duty
nomic-embed-textthe world's memory — the pinned embedder (swapping it means re-writing every vector, so it is never swapped casually)read 2026-08-11
licenceApache 2.0
what it asks of ussame §4 duties
what it permits yousame — nothing in the text claims rights over embeddings
qwen3.8:27bno seat today — it was the doorman, RuleSage's question-screening seat, from its hire on 2026-08-17 (the audit is exhibit fourteen) until 2026-08-21, when Mistral Small 3.2 took the seat; that model's row is in the new brains section belowread 2026-08-18
licenceApache 2.0 (the text packaged with the weights, read in full — 11,346 bytes, sha256 225c54bacb96eca3…)
what it asks of ussame §4 duties
what it permits yousame — no use policy, no output claims, trademarks withheld
ms-marco-MiniLM-L6-v2the careful reader — RuleSage's re-ranker, a cross-encoder trimmed to ~22 MB on the serving hostcard read 2026-08-18
licenceApache 2.0 (declared in the model card, read 2026-08-18; the deployed artifact ships only the quantized graph and carries no packaged licence text — so this row rests on the card's declaration, and says so)
what it asks of us§4 notice duties if redistributing
what it permits yourun it anywhere, commercially, royalty-free
qwen3.5:27bthe re-bake's best voice — benched, awaiting a second cardread 2026-08-11
licenceApache 2.0
what it asks of ussame §4 duties
what it permits yousame

Cards with the oxblood edge are seats our own worlds run today. Receipts from the reads, because a ledger should show its work: the 12b tag ships the Apache body without the appendix (the appendix is the how-to-apply boilerplate, not terms — noted because we read to the end, not because anything is missing); the 26b and qwen3.5:27b blobs are byte-identical stock text; qwen3-vl ships the licence twice — two byte-identical copies concatenated by the packager, most likely one per component — a 402-line dump that is still just Apache 2.0; and nomic-embed's blob is the stock text with no addendum of any kind, which is the whole basis for the no-claim-over-embeddings cell in its row. None of the five names a copyright owner in the blob — which means the §4 notice duties cannot be discharged from these files alone: the actual holders live in the upstream model cards, and ollama show --license proves what was packaged, not who licensed the weights. And the claim that matters most, stated exactly as far as our evidence reaches: the licence blobs these tags ship carry no Gemma custom terms — earlier gemma generations packaged them, these do not. That is a read of the redistributed file, not of the vendor's own terms page; when we have read that second source first-hand, this row will say so.

Bench history without a fresh stamp: gemma2:9b (the 8G-era voice pick) appears in the voice trials but is not installed on the bench box today, so we cannot re-read its shipped licence — and this ledger stamps only what we actually re-read. Its row will return if it does. The qwen3.6 of that field was in the same state on 2026-08-11; a qwen3.6 build was pulled for the August trials and read on 2026-08-12, and its row is in the next section — note the tag, qwen3.6:27b, which is not the qwen3.6:latest that field ran.

The sound department · music generation, stems, encoding · read 2026-08-25 / 26

Read from the LICENSE files in the repositories and the Hugging Face model cards themselves, in full — the same first-hand reads behind the local-music landscape study. Everything below runs on strata→signal hardware, on-prem; nothing is sent to a cloud service. The interactive tools' own sounds carry no third-party licence at all — see the closing note.

ACE-Step 1.5 (turbo · acestep-5Hz-lm)the song engine — local full-song generation; the source tracks the demos are built fromread 2026-08-25
licenceMIT (the repo LICENSE and the weights' HF cards read end-to-end — plain MIT, no acceptable-use rider; the bundled Qwen3-Embedding component is separately Apache 2.0)
what it asks of usthe MIT notice if we redistribute the weights — nothing more; its verify-originality / disclose-AI lines are guidance, not licence terms
what it permits yourun it anywhere, commercially — the card states the generated music may be used commercially. Its "licensed / royalty-free training data" line is a vendor claim with no manifest, and this row records it as a claim, not an audit
Demucs · htdemucs / htdemucs_6sthe splitter — 4- and 6-way stem separation, run on-prem on CPUread 2026-08-26
licenceMIT (© Meta Platforms; the adefossez fork LICENSE read in full, no weights carve-out)
what it asks of usthe MIT notice on redistribution
what it permits yourun it anywhere, commercially — and the separated stems carry no licence of their own; the rights that matter live in the source recording you fed it
FFmpeg (static build)the encoder — MP3 / WAV export off the generation hostinstalled 2026-08-25
licenceGPL (the johnvansickle static build bundles GPL-licensed encoders; FFmpeg's own core is LGPL-2.1+. Invoked only as a standalone binary — never linked into our code)
what it asks of usthe GPL/LGPL duties attach to distributing FFmpeg itself; running the binary to encode our own audio triggers none of them
what it permits youencode freely — the licence governs the tool, never the audio it writes

The Beat Module and the mixing tools ship no third-party audio code and no samples: every drum voice, sample-pad sound, filter sweep and reverb is synthesised from first principles in the browser's own Web Audio — nothing borrowed, nothing to license. So the sound the tools make is ours; the two models above only ever run on our own hardware, on-prem, to make the source material — nothing in this section reaches a cloud service.

The trial roster · the August arrivals and one incumbent · read 2026-08-12

Read via ollama show --license <tag> — the same method as the brains above, on the blobs pulled for the 2026-08 trials. These five rows are receipts, not readings: this cycle recorded the first line of each licence text verbatim, a sha256 of the whole text, and its size — and the row says that and stops. No verdict is pre-judged here. One of these tags carries bespoke vendor terms rather than a stock licence, and a row that summarised them from a first line would be exactly the guess this page exists to refuse. The clause-level read is queued; when it lands, these rows say more and carry a newer date. The models themselves sit the August arrivals, where their numbers are published in full.

muse-glimmer:30b-q8_0-dflashbenched on both house exams 2026-08-12 — no seatread 2026-08-12
licencethe packaged blob carries no licence text at all — a single byte, which the receipt records as 0 characters and the sha256 of the empty string (e3b0c442…). Published as a finding, not as a verdict. Read instead from the vendor's own LICENSE file on the model hub — huggingface.co/meta-models/Muse-Glimmer-30B (Meta Superintelligence Labs): Apache 2.0, read first-hand 2026-08-10 and again 2026-08-12.
what it asks of usnothing we can read out of the package — the weights ship without their licence text, which is worth knowing before anyone redistributes them
what it permits youwhat the vendor's own LICENSE file permits, read at the source — the packaged blob will not tell you, and we would rather say that than paraphrase a file that is not there
nemotron-3.5-lightning:30b-a3bbenched on both house exams 2026-08-12 — no seatread 2026-08-12
licencethe blob's first line, verbatim: "NVIDIA Open Model License Agreement" — 10 068 characters, 64 lines, sha256 aab95a9b…
what it asks of usthe clause-level read we still owe: bespoke vendor terms, not a stock licence, and this cycle read the first line and hashed the rest
what it permits youunstated here on purpose — the agreement ships inside the weights; read it there rather than take a summary from us
qwen3.6:27bbenched on both house exams 2026-08-12 — no seatread 2026-08-12
licencethe blob's first line, verbatim: "Apache License" — 11 358 characters, 203 lines, sha256 5e9ea707…
what it asks of usthe version line and the clause body were not captured by this cycle's receipt, so the row claims the first line and the sha, not a version
what it permits youread the text yourself — it ships inside the weights and prints in full from the same command named above
olmo-3.1:32b-think-q4_K_Mbenched on the judge exam 2026-08-12 — no seatread 2026-08-12
licencethe blob's first line, verbatim: "Apache License" — 11 359 characters, 203 lines, sha256 8c6db340…
what it asks of ussame as the row above: a first line and a hash are what we read, so a first line and a hash are what we claim
what it permits youread the text yourself — one character longer than its neighbour's, and a different file
granite4.1:30b-q8_0the incumbent on the roster — cleared both floors of the judge trialread 2026-08-12
licencethe blob's first line, verbatim: "Apache License" — 11 358 characters, 202 lines, sha256 cfc7749b…
what it asks of usthe same read we owe the other Apache-opening rows; it is on the roster this cycle because it sat the trials beside the new models, not because its terms moved
what it permits youread the text yourself — same opening line as the two rows above, same size as one of them (qwen3.6:27b), and still a third distinct file

Receipts, because a ledger should show its work. Three of these five blobs open with the same two words and are still three different files: 11 358 characters over 203 lines, 11 359 over 203, and 11 358 over 202, with three different sha256s between them. That is why the receipt is a hash and a size and not a name — "Apache License" is what the first line says, not proof of what the other eleven thousand characters do. The sharpest artifact of the read is the first row: a 30B model packaged with an empty licence file, which no summary of anyone's summary would ever have surfaced. The dates on these five rows are the day the text was read; what the text means is a read we still owe, and every row says so until we have made it.

The new trial roster · the candidate the 24 GB battery brought in · read 2026-08-29

One row, and it sits under its own heading rather than with the five above because it was read a different way on a different day. Laguna XS 2.1 sat the instrument travels as the one candidate of that battery with no earlier bench on this hub, and the packaged blob is not the source here: the read is the vendor's own LICENSE.md at the model repository, opened in full on 2026-08-29 and cross-read against the model card above it and the tag's own library page, which name the same licence and the same publisher.

laguna-xs-2.1:latestthe fifth candidate of the 24 GB battery, 2026-08-26 — benched, no seatread 2026-08-29
licenceOpenMDW-1.1 — the licence file's own first line, verbatim: "OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)". Published by Poolside. Three sources, read the same day and named because a maker is a fact this ledger reads rather than infers: the vendor's licence file at huggingface.co/poolside/Laguna-XS-2.1/blob/main/LICENSE.md, the model card above it, and the tag's page at ollama.com/library/laguna-xs-2.1. All three name Poolside as publisher and OpenMDW-1.1 as the licence; the model card's own sentence is "Laguna XS 2.1 is released under OpenMDW-1.1, a fully permissive license." Until this read the model roster carried no maker for the tag at all, and said so on its own page rather than guessing one.
what it asks of uskeep the agreement itself, and every copyright and origin notice included in the Model Materials, with any copy we distribute — we redistribute no weights, so nothing further attaches; the text also terminates the grant if we bring patent or copyright litigation over the materials, unless we are answering a suit brought first
what it permits youin its own words, to "deal in the Model Materials without restriction" under the copyright, patent, database and trade-secret rights embodied in them — and it "does not impose any restrictions or obligations with respect to any use, modification, or sharing of any outputs generated by using the Model Materials"

Why one row earns its own heading: the section above is stamped 2026-08-12 and its method is the packaged blob read through the runtime. This row is neither — a later day, a different source — and folding it under that heading would have made the heading untrue of one of its own rows. The other four candidates of that battery already reach the model roster through the benches that measured them.

The painters · diffusion · read 2026-08-11 except where a row says otherwise

Read via each model's authoritative record: the creator's permission flags fetched as raw JSON from civitai.com/api/v1/models/<id> (not the page prose — the ids are in the rows), and the licence files at their source repositories. The painters' bench story — who paints best — lives on the diffusion research bench; this table is only the law.

The four Civitai permission flags, in words: Image = generated images may be shared · RentCivit = the model may run on Civitai’s own generator · Rent = other hosted services may run it · Sell = the model file itself may be sold. ✓ = granted, ✗ = withheld.

albedobaseXL v2.1the painter of recordread 2026-08-11
licencethe creator's permission flags as recorded on Civitai (model 140737), over a CreativeML Open RAIL++-M SDXL base: Image ✓ · RentCivit ✓ · Rent ✓ · Sell ✗ — credit required
Two evidence bars, both ours — and this is the one row on the page that is read to both. The first bar is the row above: the creator's permission flags, read first-hand from the platform's raw record on 2026-08-11. That is the bar our campaign's own kits ship under, and nothing here withdraws it. The second bar is stricter and this painter does not clear it: our newer bench harness publishes a picture only when the licence text itself has been opened and read, and for this model that text has never been found — so the new-arrivals bench (2026-08-23, the section below) withholds this painter's images entirely. Its verdict there is unverified: its renders are counted, painter by painter, and none of its pictures is published. A permission flag is a record of what a creator ticked; a licence text is a document you can read to the end. We would rather name the gap between the two than quietly treat them as the same thing.
what it asks of uscredit the creator — given: our worlds are painted by albedobase xl, by albedobond. (A product credits page is queued and will carry it too; until it exists we say so here rather than claim it.)
what it permits youpublish and sell outputs; run it in a hosted service; may not sell the weights (Sell ✗)
sd_xl_base 1.0the sketch pipeline's baseread 2026-08-11
licenceCreativeML Open RAIL++-M (2023-07-26)
what it asks of usthe flow-down clause binds us — the use restrictions must appear as an enforceable provision in our player terms. That work is open; this row changes when it lands.
what it permits youcommercial use, hosting, sublicensing — under the licence's Attachment A use restrictions
flux.1 [schnell]the fast brush — the 4-step painter behind 2,296 of the campaign census's 2,828 renders (the census)read 2026-08-11
licenceApache 2.0 — the platform's licence record (tag: apache-2.0); the repo itself is download-gated, so this stamp covers the record. The full text was read earlier; that read's date went unrecorded — our lapse.
what it asks of us§4 notice duties if redistributing
what it permits yourun it anywhere, commercially — stock Apache 2.0, unlike its [dev] sibling two rows down
Juggernaut-XL v9bench candidate only — never wired into a productread 2026-08-11
licencethe creator's permission flags as recorded on Civitai (model 133005): Image ✓ · RentCivit ✓ · Rent ✗ · Sell ✗ — credit required
what it asks of uscredit; and the restraint itself — hosted use off civitai is not granted, so we don't
what it permits youpublish and sell outputs; do not run it inside your own hosted service
flux.1 [dev]held — shown only as research receipts (ruled 2026-09-04)read 2026-08-23
licenceFLUX.1 [dev] Non-Commercial License v1.1.1 (re-read in full 2026-08-23 for the new-arrivals bench, 18 621 bytes, alongside the v2.1 text the 2026 repacks point at — the operative sentences are the same in both, and the re-read confirmed this row rather than moving it. The same hold is enforced per cell on that bench, in the section below.)
what it asks of uswe could not satisfy ourselves, reading §2(d)'s output grant against §4(a)'s use prohibition, which one governs commercial use of dev's outputs — so we hold every flux experiment rather than guess in our own favour
what it permits youweights: non-commercial only — and §1(c) excludes any end-user-facing use even when free, §4(a) bars production as well as commercial purposes, and §2(e) conditions the output right on content filtering or pre-distribution review plus AI-generation disclosure. Outputs: §2(d)'s grant is expressly qualified ("except as expressly prohibited herein") and §4(a) survives termination (§8(d)) — on our reading, commercial use of dev outputs is not granted. We are not counsel; this is the row we most want a professional to confirm.
research receipts (ruled 2026-09-04)Nine Worlds, One Dog (published 2026-08-28) shows pictures this painter drew as receipts of a measured run — the request, the seed, the size and the seconds, all in its kit — saying so on its own face and asking that the pictures not be treated as cleared for anything else; the next gallery page painted by this painter carries the same line from its first day. Nothing of this painter's is sold or licensed onward, and no picture it drew is published on this hub outside those pages. One use we name rather than leave for a reader to find: the live world has painted 367 of its 2,828 in-play renders with this painter — 13.0%, counted in The Dog, the Dice, and the Painter — on a game that is free and sells nothing, which is the end-user-facing use §1(c) reaches even when free; that use sits under the same unresolved reading as the rest of this row, and moves when a professional's reading does. Until today this row read held — nothing published — it had said so since this ledger's first day, and it stopped being true on 2026-08-28, when nine of this painter's pictures went up on this hub as receipts; the row now says what we actually do with them. Nine Worlds said on its release day that this licence covers sharing its pictures; that was wrong on the reading above, and the page carries a dated correction since 2026-09-04
pixel-art-xl (LoRA)quarantinednot re-read this cycle
licencecreator's terms: non-commercial — verdict carried from the quarantine read, not re-read this cycle; queued
what it permits younot usable in a commercial pipeline; we pulled it from ours entirely
realvis xl 5.0 · pixel-art-redmondwithheld — unverifiednot re-read this cycle
licencenot read first-hand
what it asks of usour own rule: unread means unpublished
what it permits youunknown to us — we won't guess on your behalf
sd 3.5 · hidream · luminacandidate painters — in the census, not yet readnot re-read this cycle
licencevarious — readings queued
what it permits yourows land here as each read finishes; until then they are candidates, not recommendations

A reading trap we hit so you don't: the OpenRAIL++ file as shipped renders its Section III clauses without numerals while the text cross-references "paragraph 5" — map paragraph 4 to Distribution, 5 to Use-based restrictions, 6 to Output, or you will quote the wrong clause. And this read of the civitai flags is finer-grained than our last: the full permission arrays are now recorded verbatim (Rent present for AlbedoBase, absent for Juggernaut) — refining the prior reading, not contradicting it.

The new painters · the new-arrivals diffusion bench · all read 2026-08-23

Update, 2026-08-24 — twelve painters' rows added. What changed, in plain words: the new-arrivals diffusion bench sat thirteen open-weight image models down in front of one small world and had each of them paint it, and before a single pixel was rendered it opened every one of those painters' licences first-hand. Those reads are folded into this ledger below. The window at both ends: the licence texts were opened on 2026-08-23 (UTC); the rows landed on this page on 2026-08-24 (UTC). Against what stood here before: until today this ledger carried the three painters our game's own campaign actually ran — AlbedoBase XL 2.1, FLUX.1 [schnell] and FLUX.1 [dev] — beside the first bench's roster and its quarantined and unread LoRAs, and not one of the twelve models below had a row here at all. The painters' bench story — who paints best — lives in the diffusion wing; this table is only the law.

Read from the licence text itself, at the source the vendor actually publishes it: each file was downloaded, its byte size and sha256 recorded, and diffed against the canonical text of the licence it claims to be — so a row that says Apache 2.0 here means a file we hashed, not a tag we trusted. Each row is keyed on the checkpoint filename as it lands on disk, because a filename is the only thing a render manifest records; a quantisation does not relicense, so a repack takes its parent's reading and the row says so.

The mark at the head of each row is the verdict the bench's own publication law enforced — per painter, per cell, not per page. published means that painter's pictures are on the bench. held and unverified both mean the renders were made and are counted, painter by painter, and none of the pictures was published — and they are not the same fact: held is a licence we read that refuses republication of the outputs, unverified is a licence we have never been able to find. Every row below says which, and one row is a read with no verdict at all, because that model never painted a cell.

flux.2 [klein] 4Bpublished — the only Apache FLUXread 2026-08-23
licenceApache License 2.0 — the vendor's own LICENSE.md, beside the weights (opened 2026-08-23: 9 584 bytes, sha256 ca02bc51…; diffed against the canonical Apache text — identical modulo whitespace, bar the "how to apply" appendix, which the vendor omits. No Black Forest Labs rider, no acceptable-use annex, no output clause. The undistilled klein-base-4B ships a byte-identical file. Checkpoint: flux-2-klein-4b.safetensors. Source: huggingface.co/black-forest-labs/FLUX.2-klein-4B/blob/main/LICENSE.md)
what it asks of us§4 notice duties only if we redistribute the weights or a derivative — we ship none; §6 grants no trademark right, so we name the model and claim no endorsement
what it permits yourun it anywhere, commercially, royalty-free; publish the outputs and license them onward — no field-of-use restriction, no revenue ceiling, and nothing in the text reaches generated images. Not to be confused with the 9B in the next row, which is non-commercial: same family name, different licence
flux.2 [klein] 9B-KV (fp8)held — counts only, no pictures publishedread 2026-08-23
licenceFLUX Non-Commercial License v2.1 (opened 2026-08-23: the edit variant's LICENSE is byte-identical to Black Forest Labs' canonical v2.1 text, 18 157 bytes, sha256 e98f298d…; the distilled 9B and klein-base-9B ship the same text plus a trailing newline. A quantisation does not relicense. Checkpoint: flux-2-klein-9b-kv-fp8.safetensors. Source: huggingface.co/black-forest-labs/FLUX.2-klein-9B/blob/main/LICENSE.md)
what it asks of usthe same unresolved reading as flux.1 [dev], and we answered it the same way: §2(d) grants outputs "for any purpose (including for commercial purposes)" but only "except as expressly prohibited herein", and §4(a)(i) prohibits using "any data produced by the FLUX Model" for any commercial or production purposes. We hold rather than guess in our own favour
what it permits youweights: non-commercial only, and §1(c) excludes end-user-facing use even when free. Outputs: contested on the text's own face — and the vendor's licensing FAQ, read 2026-08-23, says images for products you charge for need a commercial licence. On our reading, a page that licenses its images onward is exactly what §4(a) forbids permitting, so this painter has no column on the bench
flux.2 [dev] (fp8mixed)held — counts only, no pictures publishedread 2026-08-23
licenceFLUX Non-Commercial License v2.1 (opened 2026-08-23 at the vendor's own code repository, which is where the canonical copy of this licence lives — 18 157 bytes, sha256 e98f298d…; the model pages carry the same text under a different filename. Checkpoint: flux2_dev_fp8mixed.safetensors. Source: github.com/black-forest-labs/flux2/blob/main/model_licenses/LICENSE-FLUX-NON-COMMERICAL)
what it asks of usthe same reading as the 9B row above — §2(d)'s output grant against §1(c) and §4(a)'s exclusions — and the same answer: the biggest model on the bench painted its whole set and none of it is published here
what it permits youweights: non-commercial only. Outputs: not something we will claim on your behalf from a text that contradicts itself and a vendor that has twice declined to resolve it in writing. Read it yourself at the link above — it is the row we would most like to be wrong about
flux.2-small-decoderan accessory, not a painter — read on the way to the bench, never wiredread 2026-08-23
licenceApache License 2.0 — declared, not shipped (the repository holds no LICENSE file, checked 2026-08-23; the claim is the repo's own metadata plus the model card's sentence "This model is licensed under the Apache 2.0 licence", and the canonical Apache text is what we opened. Checkpoint: full_encoder_small_decoder.safetensors. Source: apache.org/licenses/LICENSE-2.0, by way of huggingface.co/black-forest-labs/FLUX.2-small-decoder)
what it asks of us§4 notice duties on redistribution only. And the honest limit of this row: it carries no bench verdict, because this decoder never painted a cell and the bench's publication law holds no entry for it. A read is not a verdict, and we would rather print the gap than round it off
what it permits youwhat Apache 2.0 permits — with the structural caveat that makes this row nearly useless on its own: a decoder only replaces the last stage of the pipeline, so the base model it is paired with sets the picture's posture. Bolted to an Apache base the result stays Apache; bolted to a non-commercial base it does not become commercial
hidream-o1-image-dev (fp8)published — the pixel-space one, no VAE in the pathread 2026-08-23
licenceMIT License (opened 2026-08-23: 1 066 bytes, sha256 05660a75…, canonical MIT text with "Copyright (c) 2026 HiDream.ai" and nothing added. No LICENSE file sits in any of the model repositories — the text lives in the vendor's code repository, and the vendor's README says in as many words that the models, not only the code, are MIT. Checkpoint: hidream_o1_image_dev_fp8_scaled.safetensors. Source: github.com/HiDream-ai/HiDream-O1-Image/blob/main/LICENSE)
what it asks of uskeep the notice with any copy of the software we redistribute — we redistribute none. One thing this row does not cover: an optional prompt-rewriting stage that some published workflows wire in front of this model is a separate model under its own vendor terms, which we have not read and this row does not reach. If it is in your graph, it needs its own row
what it permits youuse "without restriction" — the licence's own words — including commercially; publish and license the outputs onward. No field-of-use clause, no acceptable-use policy, no revenue ceiling, no output clause
krea 2 turbo (fp8_scaled)published — the only row on this page with a revenue ceilingread 2026-08-23
licenceKrea 2 Community License Agreement v.1, dated 2026-06-22 (a six-page PDF, read in full 2026-08-23: sha256 b82a2805…, byte-identical at the pinned copy the model card itself designates, at the vendor's code repository, and in the ComfyUI repack of the weights. Bespoke terms, not a stock licence, which is why this row is the longest one here. Checkpoint: krea2_turbo_fp8_scaled.safetensors. Source: krea.ai/krea-2-licensing)
what it asks of usthree things, all of them real. §2.3: commercial use of the outputs is permitted only while the licensee's company-wide trailing-twelve-month revenue stays under US$1,000,000 — cross it and commercial use stops until an enterprise licence is signed, so this row gets re-read as we grow. §4.2: a content filter or human review before distribution — ours is an operator's read of every image before it ships. §4.3: disclose AI generation where law or platform policy requires; our pages name what painted every picture anyway
what it permits you§5.3 is unusually clean: "You own all Outputs you generate" — the vendor claims no ownership, and no attribution duty attaches to an image. The ceiling above binds us, the licensee; you are not a party to our agreement, so it does not travel with the pictures. The naming and notice duties in §3.1 bind anyone distributing the model or a fine-tune, which is a different act from using an image
kandinsky 5.0 image litepublished — a late addition, read before it renderedread 2026-08-23
licenceThe MIT License (MIT) (opened 2026-08-23: 1 099 bytes, sha256 94946098…, canonical MIT with "Copyright (c) 2025 Kandinsky Lab" filled in and nothing added. The weights repository carries the licence as a metadata tag only — three files in the whole tree and none of them a LICENSE. Checkpoint: kandinsky5lite_t2i.safetensors. Source: github.com/kandinskylab/kandinsky-5/blob/main/LICENSE)
what it asks of usthe MIT notice on redistributed copies of the software — we redistribute none. A correction this row also makes, since a living ledger should name its own repairs: our harness first carried this model on a guessed filename that the vendor has never published. A row keyed to a file that can never appear is a check that never runs, so the row now keys to the one file the vendor actually ships
what it permits youuse, modify and sell without field-of-use restriction; publish and license the outputs onward. No output clause, no revenue ceiling. Its text encoder and autoencoder are separately permissive, so the arm's weakest link is permissive too
z-image turbo (bf16)published — an incumbent of the first bench, re-read hereread 2026-08-23
licenceApache License 2.0 (opened 2026-08-23 at the vendor's code repository: 201 lines, md5 86d3f3a9…; diffed whitespace-insensitively against the canonical Apache text, the only difference is one leading blank line. Neither model repository carries a LICENSE or NOTICE file — the Apache claim on the model page is metadata, and the vendor's code repository holds the only text there is. If someone asks you for a licence file beside these weights, the honest answer is that there is not one. Checkpoint: z_image_turbo_bf16.safetensors. Source: github.com/Tongyi-MAI/Z-Image/blob/main/LICENSE)
what it asks of us§4 notice duties on redistribution only, and there is no NOTICE file upstream to carry in any case; §6 withholds the trade names beyond describing where the model came from
what it permits yourun it anywhere, commercially; publish and license the outputs onward. No field-of-use restriction, no acceptable-use policy we could find, no revenue ceiling, and the word "output" does not occur in the text
z-image base (undistilled)published — the slow sibling of the row aboveread 2026-08-23
licenceApache License 2.0 — the same vendor file as the turbo row (opened 2026-08-23; the same code repository's README announces both the turbo and the base release, so one text governs both. Checkpoint: z_image_bf16.safetensors. Source: github.com/Tongyi-MAI/Z-Image/blob/main/LICENSE)
what it asks of usthe same §4 duties, on redistribution we never do
what it permits youthe same — this row exists separately because the bench ran a separate checkpoint file, and a row that covers two files by assumption is a row that has stopped being read
ernie-image turbopublished — Baidu's 8-step distillread 2026-08-23
licenceApache License 2.0 — shipped beside the weights, which is rarer here than it should be (opened 2026-08-23: 11 366 bytes, sha256 5d7deeaf…; the only difference from the canonical text is the appendix placeholder filled in as "Copyright (c) 2025 Baidu, Inc. All Rights Reserved.". No rider, no acceptable-use annex, no threshold, and the word "output" does not occur. The undistilled sibling's file is byte-identical and carries the same row. Checkpoints: ernie-image-turbo.safetensors and ernie-image.safetensors. Source: huggingface.co/baidu/ERNIE-Image-Turbo/blob/main/LICENSE)
what it asks of us§4 notice duties on redistribution only; §6 withholds the trade names beyond describing origin. The diligence this row really turned on was the parts list: this pipeline bundles someone else's autoencoder and someone else's text encoder, so we checked both rather than the headline licence alone
what it permits yourun it anywhere, commercially; publish and license the outputs onward. The bundled autoencoder is the same file its own author publishes under Apache 2.0 in another model — we compared the two blobs, and their hashes match — and the bundled text encoder is Apache 2.0 in its own right, so the weakest link in this arm is still permissive
chroma1-hd (fp8mixed)published — the stylized independentread 2026-08-23
licenceApache License 2.0 — declared in prose, matched against text (no LICENSE file exists in any of the author's four model repositories, checked 2026-08-23. The declaration is the model page's metadata tag plus the README's own sentence — "It is fully Apache 2.0 licensed, ensuring that anyone can use, modify, and build upon it" — and what we opened is the canonical Apache text together with the author's own copy in their training-code repository, which matches it but for one trailing space. That is a licence read, and it is also a weaker footing than a file beside the weights; the row says so rather than levelling the two. A quantisation does not relicense, so this repack takes its parent's reading. Checkpoint: Chroma1-HD-fp8mixed.safetensors. Source: apache.org/licenses/LICENSE-2.0)
what it asks of us§4 notice duties on redistribution only. Separately, and not a licence matter: this model's training material is community-sourced and uncensored by the author's own statement, which is a content-screening duty on anyone deploying it rather than a term of the licence
what it permits yourun it anywhere, commercially; publish and license the outputs onward; fine-tune it — the author expressly invites it. No output clause, no revenue ceiling, no use restrictions in the licence
mage-flow 4B turbo (int8)published — the fastest credible sketch engine on the benchread 2026-08-23
licenceMIT License (opened 2026-08-23: 1 066 bytes, sha256 275b4dd6…, standard MIT with "Copyright (c) 2026 Microsoft"; the vendor's README carries a per-model table whose row for this model reads "MIT". Neither mirror ships a LICENSE file — the text lives only in the vendor's code repository. Checkpoint: mage_flow_turbo_int8_convrot.safetensors. Source: github.com/microsoft/Mage/blob/main/LICENSE)
what it asks of usthe MIT notice on redistributed copies — we redistribute none — and two disclosures we would rather you heard from us. One: the vendor withdrew its own model repositories at the end of 2026-07 and has published no reason, no revocation and no relicensing; the weights on our bench came from community mirrors, one of them duplicated by the paper's first author. Two: the README beside that unmodified MIT file says the models are "released for research purposes only and are not intended for product or service deployment". That is prose sitting next to a licence, not a term inside one — but it is the kind of sentence a reader deserves to see rather than discover
what it permits youuse "without restriction", including commercially; publish and license the outputs onward. No field-of-use clause in the licence, no revenue ceiling, no output clause — with the "research purposes only" note above weighed by you rather than for you

Receipts and the finding we did not expect. Across these twelve models the pattern is not the terms — most of them are stock Apache 2.0 or stock MIT, hashed and diffed against the canonical text — it is where the terms live. Z-Image, Chroma1, Mage-Flow, Kandinsky, HiDream and the small decoder all ship their weights with no licence file beside them at all: the claim is a metadata tag on a model page, and the only text that exists sits in a code repository somewhere else, or at apache.org. That is still a licence and we read it, but it is a weaker footing than a file in the same directory as the weights, and the rows say which footing each one stands on. The second finding is a family trap worth stating twice: FLUX.2 [klein] 4B is Apache 2.0 and FLUX.2 [klein] 9B is not — same vendor, same product name, one letter of difference in the size, and opposite verdicts. A ledger that read the family instead of the file would have got that backwards.

The new painters · FLUX 3 Image Klein's bench · read 2026-10-15

Update, 2026-10-20 — one painter's row added. What changed, in plain words: the FLUX 3 Image Klein bench ran Black Forest Labs' newest open-weight image model, released 2026-10-15, beside two painters this ledger already carries (flux.1 [dev] and flux.2 [klein] 4B), and that page shows the new model's pictures; this row is the licence they are shown under. The window at both ends: the licence was read on 2026-10-15 (UTC), the day the model came out; the row landed on this page on 2026-10-20 (UTC). Against what stood here before: no FLUX 3 model had a row on this page, and the newest painter row had been read on 2026-08-23.

flux 3 image klein (distilled, bf16 and int8)published — under its non-commercial use, on one bench page (ruled 2026-10-16)read 2026-10-15
licenceFLUX Kommunity License v1.0 (read 2026-10-15 from the vendor's model cards, whose License section says: "non-commercial use, plus commercial output generation by Qualifying Users with less than US$5M in gross annualized revenue, including affiliates, subject to the license terms", and whose out-of-scope section summarises the licence's Section 4. The licence text itself sits behind the repositories' download gate, which the workshop's account accepted on 2026-10-15 to fetch the weights, and it was not opened for this row: like the flux.1 [schnell] row above, this stamp covers the vendor's own record and not the file, and a full read is owed and will move this row. The text encoder the model uses, Qwen3-VL-4B-Instruct, is Apache-2.0 by the same card. Checkpoints: flux3_image_klein_sd_bf16.safetensors and flux3_image_klein_sd_int8_convrot.safetensors. Source: huggingface.co/black-forest-labs/flux-3-image-klein-comfy)
what it asks of uskeep to non-commercial use, which an evaluation is: this hub charges nothing and carries no ads. By the card's summary of Section 4, never use the model or its pictures to improve another model that does what a FLUX model does, so no picture it drew feeds a training set; and no military, surveillance or biometric use
what it permits youby the card: non-commercial use, and commercial use of its pictures for people and companies under US$5 million a year in gross revenue, affiliates included, subject to the licence's terms; anyone larger needs the vendor's commercial licence. The pictures we show are research receipts, not cleared for other use. Read the full text yourself before relying on any of this, as the card itself asks

The new brains · the doorman, a vision model and the retrieval engine's models · read 2026-10-22

Update, 2026-10-24 — seven rows added. What changed, in plain words: a weekly audit of our own, on 2026-10-19, found thirteen models our products run, or are set up to run, that had no row on this page. Seven of them are language and vision models, and their rows are here: the doorman that screens what strangers type at our doors, the last vision model our game world saw with (its vision seat has been empty since 2026-09-08), and five models inside RuleSage's retrieval engine. The window at both ends: every licence below was read on 2026-10-22 between 20:43 and 21:19 (UTC); the rows landed on this page on 2026-10-24 (UTC). Against what stood here before: none of the seven had a row, and the newest row on the page had been read on 2026-10-15. Two older rows in the first brains section changed with these: the seat lines of qwen3.8:27b and qwen3-vl:2b named seats those models no longer hold — the doorman's, which Mistral Small 3.2 has held since 2026-08-21, and the world's eyes, empty since 2026-09-08 — so both now say they hold no seat, and both lost the oxblood edge; their licences were not re-read, and their dates stand.

Two methods, each named again in its row. For the four Ollama tags: the licence layer of the build's manifest, fetched from Ollama's public registry by its sha256 — the same bytes ollama show --license prints for that build — read in full, after checking that the manifest itself hashes to the digest we hold or pin, where we have one. For the three Hugging Face models: the model card at the revision our code pins, because none of those repositories ships a LICENSE file at that revision; those rows are stamped "card read", like the MiniLM re-ranker's in the first brains section. Anyone can repeat both reads without our machines.

mistral-small3.2:24bthe doorman — reads what strangers type into the print lab, the rules desk and the Beat Lab before anything else runs, and holds a seat at the long tableread 2026-10-22
licenceApache 2.0 (the licence layer of the Ollama build we run, fetched from Ollama's public registry by its own sha256 and read in full — 11,356 bytes, sha256 43070e2d4e532684…; the build's manifest hashes to 5a408ab55df5c1b5…, the digest our server runs. It is the Apache Software Foundation's text, word for word apart from whitespace, and its copyright line is still the template's "[yyyy] [name of copyright owner]", so the file names no holder. The maker's model card, huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506 at revision 95a6d26c, declares apache-2.0, ships no LICENSE file, and names Mistral-Small-3.1-24B-Base-2503 as the base it was tuned from)
what it asks of us§4's conditions, which apply when copies are redistributed: give recipients a copy of the licence (§4(a)), mark changed files (§4(b)), keep the copyright, patent, trademark and attribution notices (§4(c)), and carry the attributions of any NOTICE file (§4(d)). We hand on no copy: the model runs on our own card
what it permits youin §2's words, "a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute" the work; and in §3 a patent licence "to make, have made, use, offer to sell, sell, import, and otherwise transfer" it, which ends for anyone who sues claiming the work infringes a patent. The text never uses the word "commercial", and it sets no field-of-use, user-count or revenue limit and no use policy; §6 grants no trademark rights
minicpm-v4.5:latestno seat today — the last model in our game world's vision seat, the world's eyes, which has been empty since this build was unloaded on 2026-09-08; the build is still on our serving machine's diskread 2026-10-22
licenceApache 2.0 (the Ollama build we hold carries no licence text at all: its manifest, which hashes to 0c40168f46d1cbf5…, the digest on our serving machine, holds only the model and its vision projector, so ollama show --license prints nothing for it. This row reads the maker instead. The model card, huggingface.co/openbmb/MiniCPM-V-4_5 at revision daef484c (its latest change, 2026-08-18), says under "Model License": "The MiniCPM-o/V model weights and code are open-sourced under the Apache-2.0 license", and links github.com/OpenBMB/MiniCPM-V/blob/main/LICENSE, which we read in full — 11,332 bytes, sha256 186c8635e1f88660…, the Apache text with "Copyright OpenBMB". This card has not always said Apache for the weights: from 2025-08-25 until a commit of 2025-09-15 (ebb1f6cb, "Update: license"), its licence section said "The usage of MiniCPM-V series model weights must strictly follow MiniCPM Model License.md" and "After filling out a 'questionnaire' for registration, MiniCPM-V 4.5 weights are also available for free commercial use". That licence, the MiniCPM Model Community License (Version 1.0, June 5, 2024), which we read in full at the revision just before its deletion (5,655 bytes, sha256 8603ca877636626b…), set terms Apache does not. Its §3, "Additional Commercial Terms", let services or products that deploy the model on "edge-side devices not exceeding 5,000 units", or applications with "a daily active user count (DAU) of less than 1 million", apply to OpenBMB and, after the registration questionnaire, "may be allowed to use it commercially for free"; anyone else had to apply to OpenBMB, "which may, at its discretion, grant permission". Its §2 required anyone who distributed or provided the model, or used any product or service of it, to display "Built with 面壁MiniCPM" prominently, and said "You may not use MiniCPM series models or their outputs and results to improve any other large language models (other than MiniCPM or its derivatives)"; its §4 and Appendix A listed restricted uses, "Any military purposes" among them. OpenBMB deleted that file, with its Chinese counterpart, from its GitHub repository two minutes before the card changed (c130da1b, 2025-09-15 03:15 UTC), in the commit that changed the LICENSE file's line from "Copyright 2024 OpenBMB" to "Copyright OpenBMB"; nothing else in that file changed. Our build is later than the change: Ollama's library page dates its last update 2026-06-02, and its latest tag is our digest. A copy fetched before 2025-09-15 was published under the earlier licence. The card also asks, in its own word "optionally", for a registration questionnaire, and adds a statement disclaiming liability for misuse; neither appears in the licence text)
what it asks of usthe same §4 conditions if we redistributed a copy, and we do not; the card's questionnaire is, in its own word, optional
what it permits youthe Apache 2.0 grant, read here from the maker's own file: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6). The device and daily-user limits, the "Built with" display, the bar on using its outputs to improve other models and the restricted-use list belonged to the earlier MiniCPM licence (see the note), not to this one
qwen3:4b-instruct-2507-q4_K_Mthe retrieval engine's built-in writer — the generator RuleSage's retrieval engine is pinned to; one of our two serving hosts overrides it with gemma4:26b and the other host's setting is unread, so this row claims no live seatread 2026-10-22
licenceApache 2.0 (two copies read in full, and they differ only in blank lines: the licence layer of the Ollama build our lock pins — 11,338 bytes, sha256 d18a5cc71b84bc4a…; the manifest hashes to 0edcdef34593eac1…, our digest — and the LICENSE file the model card links, huggingface.co/Qwen/Qwen3-4B-Instruct-2507 at revision cdbee75f — 11,343 bytes, sha256 832dd9e00a68dd83…. The five bytes between them are five blank lines; set those aside and the two are word for word the same. Both are the Apache text with one line filled in: "Copyright 2024 Alibaba Cloud")
what it asks of usthe same §4 conditions if we redistributed a copy, and we do not; here the notices to keep include the copyright line naming Alibaba Cloud
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)
qwen3-vl:8b-instructthe retrieval engine's picture captioner — describes images while documents are taken in, never while answering; no build of it is pinned, and whether any of our machines holds it is unreadread 2026-10-22
licenceApache 2.0 (this row reads the build Ollama's registry serves today, not a build we have pinned: no digest for this tag is recorded anywhere in our code, so the registry's current manifest, 70e58e37beb15464…, is the only build we can name, and it is not a receipt of ours. Its licence layer, read in full: 11,355 bytes, sha256 7339fa418c9ad3e8…, the Apache text with the template's copyright line, naming no holder. The maker's card, huggingface.co/Qwen/Qwen3-VL-8B-Instruct at revision 0c351dd0, declares apache-2.0 and ships no LICENSE file; the maker's code repository, github.com/QwenLM/Qwen3-VL, carries an Apache-2.0 LICENSE. This is not the qwen3-vl:2b of the first brains section: a different, larger model)
what it asks of usthe same §4 conditions, which apply only when copies are redistributed; we hand on no copy
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)
nomic-bert-2048 (modeling code)the code under the retrieval engine's text embedder — two Python files the embedder loads on every query, each pinned by hash; only the code runs here, none of this model's own weightscard read 2026-10-22
licenceApache 2.0 (declared in the model card, huggingface.co/nomic-ai/nomic-bert-2048 at revision 77108403, the revision our lock pins; the repository ships no LICENSE file at that revision and the card has no licence prose beyond that one line — so this row rests on the card's declaration, and says so)
what it asks of usthe same §4 conditions, only on redistribution — see the note under this section on the retrieval engine's bundle
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)
bge-reranker-v2-m3the retrieval engine's second re-ranker — loaded only when its bge mode is chosen; the default is the MiniLM re-ranker in the first brains sectioncard read 2026-10-22
licenceApache 2.0 (declared in the model card, huggingface.co/BAAI/bge-reranker-v2-m3 at revision 953dc6f6, the revision our lock pins; the repository ships no LICENSE file. Two neighbouring sources say MIT, and they cover other things: the card sends readers to its GitHub project, FlagEmbedding, whose LICENSE file is MIT and covers that project's code; and the card names its base model as bge-m3, whose own card declares MIT. This row stamps the card's apache-2.0 for these weights and names the other two so neither is mistaken for it. Our page on OpenJev as a re-ranker cites the same card as Apache 2.0, read 2026-10-10)
what it asks of usthe same §4 conditions, only on redistribution — see the note under this section on the retrieval engine's bundle
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)
nomic-embed-vision-v1.5the retrieval engine's image memory — its picture search uses it by default in code, but it is not yet in the engine's bundle or its lock, and it was not on the one serving host whose cache we have readcard read 2026-10-22
licenceApache 2.0 (declared in the model card, huggingface.co/nomic-ai/nomic-embed-vision-v1.5 at revision e3a725bc, 2025-03-31, the revision our code pins; the repository ships no LICENSE file. This card has not always said Apache: it declared cc-by-nc-4.0, a non-commercial licence, until a commit of 2025-01-16 (351ebb2e) changed that line to apache-2.0; no other words on the card changed. We read the card on both sides of that commit, at af2246ff (2024-06-08) and at 351ebb2e. Our pin is later than the change; a copy fetched before 2025-01-16 was published under the earlier declaration)
what it asks of usthe same §4 conditions, only on redistribution; the engine has not fetched it into its bundle, so nothing attaches today
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)

Receipts, because a ledger should show its work. Three of the four Ollama manifests hash to digests of ours (Mistral Small 3.2's 5a408ab5…, which our server runs, MiniCPM-V 4.5's 0c40168f…, which it holds, and Qwen3 4B's 0edcdef3…, which the retrieval engine's lock pins); the fourth, Qwen3-VL 8B's, has no digest of ours to compare. The three licence layers present are all the Apache Software Foundation's text apart from whitespace, and only Qwen3 4B's names a copyright holder, so for the other two the holder lives in the maker's card, not the file. None of these copies leaves our machines. The retrieval engine's own build gathers its Hugging Face and OCR files into a bundle for our own two serving hosts; if that bundle were ever handed to anyone else, each licence's redistribution conditions would travel with it.

The sound department · a voice · card read 2026-10-22

Update, 2026-10-24 — the voice's row added. What changed, in plain words: the same 2026-10-19 audit found that Kokoro, the speech model that reads aloud for the long table and the Beat Lab, had no row; until now this department held only music models and an encoder. The window at both ends: read on 2026-10-22 between 20:43 and 21:03 (UTC); the row landed on 2026-10-24 (UTC). Against what stood here before: no speech model had a row on this page, and this department's newest row had been read on 2026-08-26. One difference from the music rows above: the voice does not run on our own hardware on-prem but on the ordinary processors of our two rented web servers, so that section's "on-prem" lines speak of its own rows, not of this one.

Read from the maker's model card and, in full, the LICENSE files of the two projects that converted and packaged the model, plus the PyPI page of the package that runs it. The weights' licence rests on the card, because the maker's repository ships no LICENSE file, so the row is stamped "card read".

Kokoro-82M v1.0 (kokoro-v1.0.onnx + voices-v1.0.bin)the voice — reads aloud for the long table and the Beat Lab, on our web servers' ordinary processors rather than a graphics cardcard read 2026-10-22
licenceApache 2.0 (the weights) · MIT (the ONNX conversion and its packaging) (three hands, each read. The weights' maker, hexgrad, declares apache-2.0 on the card huggingface.co/hexgrad/Kokoro-82M at revision f3ff3571, which ships no LICENSE file and says "With Apache-licensed weights, Kokoro can be deployed anywhere from production environments to personal projects"; so the weights' row rests on the card. The ONNX graph we run was converted by github.com/taylorchu/kokoro-onnx (release v0.2.0; its LICENSE, read in full, is MIT, © 2025 taylorchu). The two files we download come from github.com/thewh1teagle/kokoro-onnx's release model-files-v1.0, whose asset sizes equal our pinned files' (325,532,387 and 28,214,398 bytes; GitHub publishes no hash for them); that repository's LICENSE, read in full, is MIT (© 2025 github.com/thewh1teagle), and the package's PyPI page for kokoro-onnx 0.4.7 states "kokoro-onnx: MIT" and "kokoro model: Apache 2.0". Each MIT file speaks of "this software and associated documentation files" and neither names the weights. The card also lists two Creative Commons Attribution sources in Kokoro's training audio, Koniwa (CC BY 3.0) and SIWIS (CC BY 4.0). One tool question this row does not settle: the package that runs the voice, kokoro-onnx 0.4.7, imports phonemizer-fork 3.3.1, whose PyPI record declares GPL-3.0-or-later, and espeakng-loader 0.2.4, an MIT package, and hands phonemizer the espeak-ng shared library to load, which espeak-ng's own repository licenses GPL-3.0. That code loads into the voice's own process; it is not a separate program like the FFmpeg row's, so how those licences attach is a question for the tool rows this ledger still owes)
what it asks of usApache §4's conditions for the weights and MIT's one condition for the conversions ("The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software"), both only when copies are handed on; we hand on no copy: the voice runs on our own servers
what it permits youfor the weights, the Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell". For the conversions, MIT's "to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies". Neither text sets a field-of-use, user-count or revenue limit, and neither says anything about the speech the model makes

The readers · layout, tables and OCR when documents are taken in · card read 2026-10-22

Update, 2026-10-24 — five rows added, under a new heading. What changed, in plain words: the same 2026-10-19 audit found five models RuleSage's retrieval engine runs only while it takes documents in — one that finds the parts of a page, one that reads the structure of tables, and three small OCR models that find lines of printed text, check which way up they are, and read them. None of them is a language model, so they have their own heading rather than a place among the brains. The window at both ends: read on 2026-10-22 between 20:43 and 21:19 (UTC); the rows landed on 2026-10-24 (UTC). Against what stood here before: none of the five had a row, and this heading did not exist.

Each of these five is declared on the page its file comes from, and none of those pages ships a LICENSE file beside the weights, so every row is stamped "card read": the Hugging Face card at the revision our lock pins for the two Docling models, and RapidAI's model repository on ModelScope at its v3.9.0 revision for the three OCR files, matched to the hashes our lock pins through that repository's own file listing. The licence texts those declarations name were read in full.

docling-layout-heronthe page reader — finds the titles, tables, pictures and captions on each page when documents are taken in, only when that richer intake is switched oncard read 2026-10-22
licenceApache 2.0 (declared in the model card, huggingface.co/docling-project/docling-layout-heron at revision 8f39ad3c, the revision our lock pins; the repository ships no LICENSE file and the card has no licence prose beyond that one line, so this row rests on the declaration and says so)
what it asks of us§4's conditions, which apply when copies are redistributed: give recipients a copy of the licence (§4(a)), mark changed files (§4(b)), keep the copyright, patent, trademark and attribution notices (§4(c)), and carry the attributions of any NOTICE file (§4(d)). We hand on no copy: the retrieval engine runs it on our own servers (see the note under this section)
what it permits youin §2's words, "a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute" the work; and in §3 a patent licence "to make, have made, use, offer to sell, sell, import, and otherwise transfer" it, which ends for anyone who sues claiming the work infringes a patent. The text never uses the word "commercial", and it sets no field-of-use, user-count or revenue limit and no use policy; §6 grants no trademark rights
docling-models (TableFormer accurate + fast)the table reader — works out the rows and columns of each table the page reader found, only when the richer intake is switched oncard read 2026-10-22
licenceCDLA-Permissive-2.0 or Apache 2.0 (dual-licensed by the maker since 2025-12-03) (the Community Data License Agreement – Permissive, Version 2.0, declared in the model card, huggingface.co/docling-project/docling-models at revision fc0f2d45, the revision our lock pins; no LICENSE file in the repository and no licence prose on the card. The card has changed since our pin: on 2025-12-03 the maker merged a change titled "Dual licensing" (21993208, whose parent is our pin), its maintainer writing "this is the simplest way for us to show the dual licensing", and the card now declares both cdla-permissive-2.0 and apache-2.0. The two TableFormer weight files are byte-identical at both revisions (Hugging Face's sha256 hashes 2a7d6c924b3cd12f… and 3119563aab5a7c96…), so the files we run are today also offered under Apache 2.0, with its patent licence (§3). CDLA-Permissive-2.0 is a licence written for data, not for models, and its text, read in full at cdla.dev/permissive-2-0, says so in its own definitions: §5.1 makes "Data" "the material received by a Data Recipient under this agreement" — here, the weights — and §5.4 counts "machine learning models" among the "Results" of analysing Data. It has no patent clause and no trademark clause. The code that runs these weights is a separate thing under a separate licence: docling-ibm-models 3.13.3, MIT by its LICENSE at the v3.13.3 tag (© 2024 International Business Machines), read in full. The card's licence for the weights and the code's LICENSE name different licences, and this row names both)
what it asks of us§2.1, its one condition, verbatim: "A Data Recipient may share Data, with or without modifications, so long as the Data Recipient makes available the text of this agreement with the shared Data." We share no copy; if the retrieval engine's bundle is ever handed on, the agreement's text goes with these files. For the code, MIT's notice on copies
what it permits you§1.1: "A Data Recipient may use, modify, and share the Data made available by Data Provider(s) under this agreement if that Data Recipient follows the terms of this agreement"; §3.1: the agreement "does not impose any restriction or obligations with respect to the use, modification, or sharing of Results". The text does not mention commercial use either way and sets no field-of-use, user-count or revenue limit. Under the Apache 2.0 offer the card has carried since 2025-12-03, the same grant as the page reader's row above: §3's patent licence reaches "use, offer to sell, sell"; no trademark rights (§6)
PP-OCRv6_det_small.onnxthe text finder — marks where the lines of text are on a scanned page, only when the OCR intake is switched oncard read 2026-10-22
licenceApache 2.0 (the file's own home is RapidAI's model repository on ModelScope, modelscope.cn/models/RapidAI/RapidOCR: the rapidocr 3.9.0 package's own model table names its download as that repository's onnx/PP-OCRv6/det/PP-OCRv6_det_small.onnx at the v3.9.0 revision, and the repository's file listing gives that file the sha256 our lock pins, 090f04abcd9d9a74…. The repository declares "Apache License 2.0" in its card and its metadata and ships no LICENSE file, so this row rests on that declaration. Two more Apache texts stand behind it, both read in full: the RapidOCR project's LICENSE at its v3.9.0 tag (11,422 bytes, sha256 3e0af25fdd06aa95…, which opens "Copyright (c) 2021 RapidOCR Authors" and gives "Copyright (c) 2021 RapidAI" in its appendix) and that of PaddleOCR, the project that trained the PP-OCR models (11,376 bytes, sha256 3840c5c0c61c2942…, "Copyright (c) 2016 PaddlePaddle Authors"). PaddlePaddle's own ONNX export of this model on Hugging Face is also declared apache-2.0, but it hashes differently, so it is not the file we run)
what it asks of usthe same §4 conditions, only on redistribution — see the note under this section on the retrieval engine's bundle
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)
PP-OCRv6_rec_small.onnxthe text reader — turns each line the text finder marked into characterscard read 2026-10-22
licenceApache 2.0 (the file's own home is RapidAI's model repository on ModelScope, modelscope.cn/models/RapidAI/RapidOCR: the rapidocr 3.9.0 package's own model table names its download as that repository's onnx/PP-OCRv6/rec/PP-OCRv6_rec_small.onnx at the v3.9.0 revision, and the repository's file listing gives that file the sha256 our lock pins, 6f327246b50388f3…. The repository declares "Apache License 2.0" in its card and its metadata and ships no LICENSE file, so this row rests on that declaration. Two more Apache texts stand behind it, both read in full: the RapidOCR project's LICENSE at its v3.9.0 tag (11,422 bytes, sha256 3e0af25fdd06aa95…, which opens "Copyright (c) 2021 RapidOCR Authors" and gives "Copyright (c) 2021 RapidAI" in its appendix) and that of PaddleOCR, the project that trained the PP-OCR models (11,376 bytes, sha256 3840c5c0c61c2942…, "Copyright (c) 2016 PaddlePaddle Authors"). As with the text finder, PaddlePaddle's own ONNX export on Hugging Face hashes differently)
what it asks of usthe same §4 conditions, only on redistribution — see the note under this section on the retrieval engine's bundle
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)
ch_ppocr_mobile_v2.0_cls_mobile.onnxthe orientation check — tells whether a line of text is upside down before the text reader reads itcard read 2026-10-22
licenceApache 2.0 (the file's own home is RapidAI's model repository on ModelScope, modelscope.cn/models/RapidAI/RapidOCR: the rapidocr 3.9.0 package's own model table names its download as that repository's onnx/PP-OCRv4/cls/ch_ppocr_mobile_v2.0_cls_mobile.onnx at the v3.9.0 revision, and the repository's file listing gives that file the sha256 our lock pins, e47acedf663230f8…. The repository declares "Apache License 2.0" in its card and its metadata and ships no LICENSE file, so this row rests on that declaration. Two more Apache texts stand behind it, both read in full: the RapidOCR project's LICENSE at its v3.9.0 tag (11,422 bytes, sha256 3e0af25fdd06aa95…, which opens "Copyright (c) 2021 RapidOCR Authors" and gives "Copyright (c) 2021 RapidAI" in its appendix) and that of PaddleOCR, the project that trained the PP-OCR models (11,376 bytes, sha256 3840c5c0c61c2942…, "Copyright (c) 2016 PaddlePaddle Authors").)
what it asks of usthe same §4 conditions, only on redistribution — see the note under this section on the retrieval engine's bundle
what it permits youthe same Apache 2.0 grant: §3's patent licence reaches "use, offer to sell, sell"; no field-of-use, user-count or revenue limit; no trademark rights (§6)

As with the brains above, none of these files leaves our machines: the retrieval engine's bundle is built for our own two serving hosts, and if it were ever handed on, the CDLA's text and the Apache conditions would go with it.

The new painters · the trial roster's one painter, Qwen-Image-2.1-Turbo · read 2026-10-22

Update, 2026-10-24 — one benched painter's row added. What changed, in plain words: the same 2026-10-19 audit found that Qwen-Image-2.1-Turbo, the picture model one of our bench pages ran once, on 2026-10-15, had no row here, although that page quotes its licence and prints its receipt. It is not in service: the copy was removed after the run and it holds no seat, so its row is a trial-roster row, filed with the painters because it is a picture model. The window at both ends: the licence was read on 2026-10-22 at 20:50 (UTC); the row landed on 2026-10-24 (UTC). Against what stood here before: no Qwen-Image model had a row on this page, and the newest painter row had been read on 2026-10-15.

Read from the LICENSE file in the maker's repository on Hugging Face, at the revision the bench ran, in full; its size and sha256 equal the receipt that bench page prints. The card's front matter was read through Hugging Face's API at the same revision.

qwen-image-2.1-turboon the trial roster — benched once, 2026-10-15, for one evaluation page, which does not show its pictures; the copy was removed after the run, so it is not in service and holds no seatread 2026-10-22
licenceQwen RESEARCH LICENSE AGREEMENT (its own name, as its first line spells it; its third line adds "Release Date: September 20, 2026". The LICENSE file at huggingface.co/Qwen/Qwen-Image-2.1-Turbo revision d65dbc9a, read in full — 7,831 bytes, sha256 8dc973f024ff9596…, the same bytes that page's receipt names. The licensor, by its §1(b): Hangzhou Tongyi Laboratory Technology Co., Ltd. The card's front matter says license: other, license_name: qwen-research)
what it asks of us§1(i): "'Non-Commercial' shall mean for research or evaluation purposes only", which an evaluation on a hub that charges nothing is. If we handed a copy on, §3's conditions would attach: a copy of the agreement to every recipient (§3(a)), notices on changed files (§3(b)) and its attribution line in a "Notice" file (§3(c)); we hand on nothing, and our copy was removed after the run. §4(b) asks that a model made, trained, fine-tuned or improved with it or its outputs, if distributed or made available, display "Built with Qwen" or "Improved using Qwen"; nothing it made went into the game or any other tool, as that page says
what it permits you§2(a): a licence "to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Materials FOR NON-COMMERCIAL PURPOSES ONLY"; §2(b): "You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us." §5(c) ends every licence it grants to anyone who brings a suit or other proceeding claiming the Materials or their output infringe a right that person owns or can license. No clause says who owns a picture it makes

Why this page is called living

A licence table with no date is a rumor with a straight face. Licences move — new versions, changed permission flags, revised clauses (flux's licence is at v1.1.1 and now covers a family of variants it didn't at our first read). So this ledger is re-read, not archived: every row carries the date its licence text was last read in full, we re-read whenever a seat changes or a model updates, and when a clause moves the row moves and the change is named in plain text. This page is the licensing authority for everything we publish — the deep dive, the diffusion bench, and the voice trials carry short credits and summaries, but none of them is the authority: where a page's wording and this ledger's dated row differ, the row wins, and the difference is a bug we want reported. Two things we owe you about the word living. The triggers above are ours to notice, so we also set a floor: every row gets a fresh read at least every six months. If the newest date on this page is older than that, we have slipped — that is a bug, and telling us is the fastest fix.

Provenance

  • First-hand — every dated row was read from the licence text itself: the file packaged with the weights, the platform's raw permission record, or the vendor's current licence file. The method is named per table so you can repeat the read.
  • Dated — the stamp is the day (UTC) of the last full read, not the day the row was written. No date, no claim.
  • Refusals count too — unread models are withheld and say so. The diffusion bench's 1,261 considered renders split 661 published / 338 held / 262 withheld, and 84 of the 262 trace to three models across the two rows above: realvis xl 5.0 (48) and pixel-art-redmond (18), both unread, and pixel-art-xl (18), quarantined. The flux hold is enforced at the experiment level — every flux experiment is held whole — rather than per-render in the publication law; folding it into the law's own reason codes is queued work, and we say so instead of rounding it off.
  • Authorship — read, drafted, and audited by the workshop's own agents under a human operator's rulings, then revised with that operator — often across many rounds; nothing releases until they have read it and signed off. The same division of labor this whole hub practices — told in full here.

We wrote this page because we got tired of guessing, and because someone deciding whether to trust a small studio's software shouldn't have to read a shelf of licences to do it. So we read them, wrote down what we found, and put the dates on so you can tell how fresh the reading is — including the rows where the honest entry is a dash. The shelf grows; the ledger grows with it.

If something here helped you, or you use our research or our data kits in your own work, please link back to this page.

elsewhere in the workshop

a strata→signal property · hello@strata2signal.com · say hello