{ "schema": "s2s-bench-v1", "dataset": "vram-residency-census", "exhibit": "august-arrivals", "table": "residency", "published_utc": "2026-08-12", "status": "PUBLISHED — live at research.strata2signal.com.", "licence": "CC BY 4.0", "attribution": "strata→signal research, research.strata2signal.com", "what_this_is": "What each model on our roster actually costs to keep resident, measured one arm at a time on the workstation that serves our own apps. A row is one (model, context length) pair, because size_vram is weights plus KV cache and the KV cache is a function of the context length — so a residency figure quoted without its context length is not a figure. These are allocations this daemon made under this configuration, not a claim about what a model needs anywhere else.", "columns": [ "model", "params_and_quant", "ctx", "loaded_vram_gib", "load_s", "note" ], "columns_note": "The six fields above are the page table's six columns, verbatim; every other field on a row is the receipt behind one of them. `size_vram_bytes` is the published figure — `loaded_vram_gib` is derived from it by the counting rule and never the other way round. A null is a number we do not hold; the page renders it as an em dash, never as a zero.", "hardware": "One 96G VRAM workstation, which also answers live requests for three of our apps (two public, one serving playtests). Two models on this roster were resident and serving users throughout, and are read rather than loaded for exactly that reason.", "daemon": { "ollama_version": "0.32.9", "kv_cache_type": "q8_0", "kv_cache_status": "MEASURED — read from the serving daemon's own unit environment at run time, not assumed. The census record names the host it was read on; that name is withheld here by the same house rule that keeps filesystem paths off this site.", "flash_attention": "1", "note": "Read from the daemon's own unit environment at run time, not assumed. Flash attention and a q8_0 KV cache are in force; a different daemon configuration would allocate differently, which is why the configuration is published beside the figures." }, "counting_rules": { "registered_rule": "size_vram is read verbatim from ollama's /api/ps for the named tag, in bytes, immediately after a load that generated exactly 1 token at the named num_ctx, with the model still resident and no other candidate resident. GB figures anywhere downstream are size_vram / 1073741824 (GiB, 1024^3), rounded to 2 dp, and the byte count always travels with them. A row is one (tag, num_ctx) pair: size_vram = weights + KV cache, and the KV cache is a function of num_ctx, so a figure quoted without its ctx is not a figure. Protected residents (the live seat, the live embedder) are READ, never loaded: their rows carry the ctx the daemon reports, which is whatever production last set, not a value we chose.", "row_is_a_pair": "A row is one (model, num_ctx) pair. The four rows at 16 384 are the four August candidates at the context length the seat trial's frozen exam actually ran; every loadable arm also carries a row at 32 768, the standard our own services pin.", "gib": "loaded_vram_gib = size_vram_bytes / 1073741824, rounded to 2 dp. The bytes travel with it.", "load_s": "load_s is the daemon's own load_duration counter in seconds, recorded as a secondary. It is a WARM page-cache load and is not comparable to a vendor's first-load figure or to a cold boot.", "ctx_confirmed": "A row's context length is confirmed when /api/ps reports the length the row asked for. A row that got a different one is about a different measurement and is not published under ours." }, "outcome_vocabulary": { "MEASURED": "loaded alone, read, context length confirmed, released. A figure is published.", "READ-ONLY": "a protected production resident, read from /api/ps while production held it. A figure is published, with the context length production set.", "NOT-OBSERVED": "a protected resident that was not resident at read time. No figure, and the reason rides the row.", "NOT-RUN": "a live workload this census may not load, release or sweep, and which was not resident to be read. No figure, and the reason rides the row." }, "outcomes": { "MEASURED": 14, "READ-ONLY": 1, "NOT-RUN": 1, "NOT-OBSERVED": 1 }, "rows": [ { "model": "muse-glimmer:30b-q8_0-dflash", "params_and_quant": "27.9B Q8_0", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 16384, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "2.31", "size_vram_bytes": 2481591417, "load_s": 16.242, "outcome": "MEASURED", "protected": false, "class": "dense 27.9B q8 + drafter", "note": "drafter pair — figure under review, see below", "reason": null, "measured_at": "2026-08-12T18:19:06-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "census-exam" }, { "model": "muse-glimmer:30b-q8_0-dflash", "params_and_quant": "27.9B Q8_0", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "2.34", "size_vram_bytes": 2515145849, "load_s": 7.713, "outcome": "MEASURED", "protected": false, "class": "dense 27.9B q8 + drafter", "note": "drafter pair — figure under review, see below", "reason": null, "measured_at": "2026-08-12T18:19:23-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "muse-glimmer:30b-q4_K_M-dflash", "params_and_quant": "27.9B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "2.34", "size_vram_bytes": 2515145849, "load_s": 13.537, "outcome": "MEASURED", "protected": false, "class": "dense 27.9B q4 + drafter", "note": "drafter pair — figure under review, see below", "reason": null, "measured_at": "2026-08-12T18:19:32-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "muse-glimmer:30b", "params_and_quant": "27.9B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "15.60", "size_vram_bytes": 16750089337, "load_s": 3.662, "outcome": "MEASURED", "protected": false, "class": "dense 27.9B q4, no drafter", "note": "the same weights, no drafter", "reason": null, "measured_at": "2026-08-12T18:19:47-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "gemma4:26b", "params_and_quant": "25.8B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "as-observed — production set this, not the census", "ctx_confirmed": null, "loaded_vram_gib": "16.27", "size_vram_bytes": 17474319810, "load_s": null, "outcome": "READ-ONLY", "protected": true, "class": "MoE 25.8B total / 3.8B active — THE LIVE SEAT", "note": "read as found — the live seat; MoE, 3.8B active", "reason": "read from /api/ps while production held it; never loaded, never released, never re-pinned by this census", "measured_at": "2026-08-12T18:19:52-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "read-only (/api/ps)" }, { "model": "gemma4:12b", "params_and_quant": "11.9B q4", "params_and_quant_source": "roster-labelled — this row was never read, so no first-hand metadata exists", "ctx": null, "ctx_source": "as-observed — production set this, not the census", "ctx_confirmed": null, "loaded_vram_gib": null, "size_vram_bytes": null, "load_s": null, "outcome": "NOT-RUN", "protected": true, "class": "dense 11.9B q4 — LIVE WORKLOAD (interim ruling 2026-08-12)", "note": "NOT-RUN — a live workload no trial may load or sweep; not resident at census time", "reason": "gemma4:12b was not resident at census time. Under the 2026-08-12 interim ruling it is a live workload this leg may not load, release or sweep, so there is nothing to read and no number is published.", "measured_at": "2026-08-12T18:19:52-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "read-only (/api/ps)" }, { "model": "gemma4:12b-it-q8_0", "params_and_quant": "11.9B Q8_0", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "12.48", "size_vram_bytes": 13398316153, "load_s": 9.166, "outcome": "MEASURED", "protected": false, "class": "dense 11.9B q8 (Leg B anchor)", "note": "", "reason": null, "measured_at": "2026-08-12T18:19:53-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "qwen3.5:27b", "params_and_quant": "27.8B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "16.24", "size_vram_bytes": 17434966751, "load_s": 9.045, "outcome": "MEASURED", "protected": false, "class": "dense 27.8B q4", "note": "", "reason": null, "measured_at": "2026-08-12T18:20:03-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "qwen3.6:27b", "params_and_quant": "27.8B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 16384, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "15.63", "size_vram_bytes": 16784849631, "load_s": 9.044, "outcome": "MEASURED", "protected": false, "class": "dense 27.8B q4", "note": "", "reason": null, "measured_at": "2026-08-12T18:20:13-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "census-exam" }, { "model": "qwen3.6:27b", "params_and_quant": "27.8B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "16.24", "size_vram_bytes": 17434966751, "load_s": 9.025, "outcome": "MEASURED", "protected": false, "class": "dense 27.8B q4", "note": "", "reason": null, "measured_at": "2026-08-12T18:20:26-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "nemotron-3.5-lightning:30b-a3b", "params_and_quant": "32.9B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 16384, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "23.60", "size_vram_bytes": 25336532172, "load_s": 11.347, "outcome": "MEASURED", "protected": false, "class": "MoE 32.9B total / 3B active q4", "note": "MoE — 3B active of 32.9B", "reason": null, "measured_at": "2026-08-12T18:20:36-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "census-exam" }, { "model": "nemotron-3.5-lightning:30b-a3b", "params_and_quant": "32.9B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "23.64", "size_vram_bytes": 25386863820, "load_s": 11.373, "outcome": "MEASURED", "protected": false, "class": "MoE 32.9B total / 3B active q4", "note": "MoE — 3B active of 32.9B", "reason": null, "measured_at": "2026-08-12T18:20:48-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "granite4.1:30b-q8_0", "params_and_quant": "28.9B Q8_0", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "33.30", "size_vram_bytes": 35754994564, "load_s": 12.731, "outcome": "MEASURED", "protected": false, "class": "28.9B q8 (seated chair-holder)", "note": "a seated chair-holder", "reason": null, "measured_at": "2026-08-12T18:21:01-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "nemotron3:33b", "params_and_quant": "33.0B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "23.21", "size_vram_bytes": 24922428537, "load_s": 11.907, "outcome": "MEASURED", "protected": false, "class": "33.0B q4 (seated chair-holder)", "note": "a seated chair-holder", "reason": null, "measured_at": "2026-08-12T18:21:14-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "olmo-3.1:32b-think-q4_K_M", "params_and_quant": "32.2B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 16384, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "18.79", "size_vram_bytes": 20171320196, "load_s": 8.465, "outcome": "MEASURED", "protected": false, "class": "dense 32.2B q4 (64K-ctx model)", "note": "", "reason": null, "measured_at": "2026-08-12T18:21:27-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "census-exam" }, { "model": "olmo-3.1:32b-think-q4_K_M", "params_and_quant": "32.2B Q4_K_M", "params_and_quant_source": "read first-hand from the /api/ps entry", "ctx": 32768, "ctx_source": "registered by this census", "ctx_confirmed": true, "loaded_vram_gib": "18.82", "size_vram_bytes": 20204874628, "load_s": 8.471, "outcome": "MEASURED", "protected": false, "class": "dense 32.2B q4 (64K-ctx model)", "note": "", "reason": null, "measured_at": "2026-08-12T18:21:37-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "house-law" }, { "model": "nomic-embed-text:latest", "params_and_quant": "137M F16", "params_and_quant_source": "roster-labelled — this row was never read, so no first-hand metadata exists", "ctx": null, "ctx_source": "as-observed — production set this, not the census", "ctx_confirmed": null, "loaded_vram_gib": null, "size_vram_bytes": null, "load_s": null, "outcome": "NOT-OBSERVED", "protected": true, "class": "137M F16 embedder — THE LIVE EMBEDDER", "note": "NOT-OBSERVED — the live embedder, protected; not resident at read time, and loading it for a number is out of contract", "reason": "nomic-embed-text:latest was not resident at read time. It is a protected production resident: loading it to obtain a number is out of contract, so this row carries no number.", "measured_at": "2026-08-12T18:21:46-06:00", "kv_cache": "q8_0", "ollama_version": "0.32.9", "client": "read-only (/api/ps)" } ], "n_rows": 17, "n_registered": 17, "complete": true, "limits": [ "size_vram is what this daemon allocated under this configuration — not what a model 'needs', and not what it would allocate under a different KV cache type, a different attention implementation or a different runtime.", "Two arms at the same size are not the same model. Quantization is published on every row; comparing across quantizations is a confound we name rather than one we correct for.", "The three drafter rows are under review and marked as such: see `open_findings`. They are printed rather than dropped, because a figure withdrawn silently is worse than a figure published with its doubt attached.", "Nothing here is a benchmark. It is an inventory of what our own roster costs to hold resident, on our own hardware, for our own chairs." ], "open_findings": [ { "id": "drafter-rows-under-review", "severity": "unresolved at publication", "rows": [ "muse-glimmer:30b-q8_0-dflash @ 16384", "muse-glimmer:30b-q8_0-dflash @ 32768", "muse-glimmer:30b-q4_K_M-dflash @ 32768" ], "what": "The three tags served with a draft model beside them report 2.31–2.34 GiB, against 15.60 GiB for the same 27.9B weights served without a drafter — and the q8_0 and q4_K_M drafter tags report a byte-identical 2 515 145 849 at 32 768 despite different digests and different quantization levels, which two different quantizations cannot both allocate.", "why_the_run_did_not_catch_it": "The census refuses a row whose reported context length differs from the one it registered, and every one of these three confirmed its context length. A reading that is internally consistent but describes the wrong entry is not a condition the registered falsification rule tests for.", "status": "published verbatim, flagged, and held pending a re-measure. No number here has been adjusted." } ], "provenance": { "artifacts": [ { "role": "the census record — every figure on this page is a copy out of this file", "artifact": "vram-census.json", "sha256": "8c53028e8b22a6e1b76c70aa7c208939f3f1e0f8974eda852db2920a55ebf130" }, { "role": "the registered denominator — the rows this census was allowed to measure", "artifact": "c7-roster.json", "sha256": "d6f67db93acf55f10b941aab3a64ee46d97df2881bb032891daac779be586844", "registered_rows": 17 }, { "role": "pre-registration, frozen before the first load", "artifact": "PREREG-C7-census.md", "sha256": "3ea13e0676300032c6ca31849c84eace65108e74b1f11be3b49ed6123350afc0" }, { "role": "the census itself", "artifact": "vram_census.py", "sha256": "0c87af2760c790f0fd8b71e3da7d77845c669e414863466508fd5b1ba378bcde" } ], "measured": "2026-08-12, evening, local time", "reading": "size_vram is read verbatim off the daemon's /api/ps for the named tag, with the whole entry kept beside it — digest, context length, expiry and vendor metadata — so every derived figure can be recomputed from the record rather than trusted.", "first_hand": "params and quantization on fifteen of seventeen rows are read first-hand from the /api/ps entry the figure came out of. The two rows the census never read carry the frozen roster's own label, and say so on the row.", "no_paths": "Artifacts are named, never located. No box names, no addresses, no filesystem paths." }, "contamination_caveat": "Published 2026-08-12. These are residency measurements, not a scored fixture, so training-set contamination does not apply to them — but the same posture does: our own trials, our own hardware, for our own chairs. Not a standard.", "not_a_standard": "Our own trials, our own hardware, for our own chairs. Not first independent numbers, and not a benchmark." }