# A short history of llama.cpp

*One evening's hack in March 2023 that put a large language model on a laptop's processor, the small C library and the company behind it, the file format it gave the world, and the servers, apps and engines that run on it today, dated from the project's own record.*

*Published 2026-11-14 (UTC) · A small (human) team and a fleet of AI agents.*

**the short version:** llama.cpp is free, open-source software that runs a large language model, the kind of AI inside a chatbot, on an ordinary computer: a laptop's processor, a graphics card, a phone. Georgi Gerganov wrote the first version in what his README called an evening and published it on 2023-03-10, two weeks after Meta released its LLaMA models to researchers. It is the engine under Ollama, LM Studio and the Strata engine, the project that wrote GGUF, the file format most of Ollama's library ships in, and the program beneath most of the language-model numbers on this hub. (Despite the similar name, the Strata engine, github.com/Niko1221/Strata, is an unrelated project; strata2signal has no affiliation with it.) Its company, ggml.ai, joined Hugging Face on 2026-02-20, and its first major version, v1.0.0, was published on 2026-11-13. It sits beside our other short histories, among them [Mistral](https://research.strata2signal.com/a-short-history-of-mistral/), [Ollama](https://research.strata2signal.com/a-short-history-of-ollama/) and [vLLM](https://research.strata2signal.com/a-short-history-of-vllm/).

3,461 words · about 16 minutes (at 220 words/min) · 1 table · data kit: no

https://research.strata2signal.com/a-short-history-of-llama-cpp/

---

*If you're new here: [strata→signal](https://strata2signal.com) is a small workshop (plus a friendly dog with a white patch) that builds on its own machines and writes up what it measures. Every date below was read during a pass on 2026-10-14, 2026-10-29 or 2026-11-12, the newest on 2026-11-13 (UTC) with the project's releases, each from the source credited beside it — among them the project's own commits, pull requests, releases and announcements as GitHub records them, the ggml.ai and llama.app pages, Hugging Face's documentation and blog, Meta's announcement, Hacker News, and our own released pages. Where we found no source, the claim is not here.*

*Five words this page leans on.* A model's **weights** are the numbers it learned in training, stored as files. A **token** is the unit a model reads and writes: a word or a piece of one. **Quantization** is storing the weights in fewer bits than they were trained in, so a model fits in less memory; [The compressed photograph](https://research.strata2signal.com/the-compressed-photograph/) explains what that costs. **GGUF** is llama.cpp's single-file format for a model, usually a quantized one. A **backend** is the code that does the arithmetic on one kind of hardware: a processor, an NVIDIA card, an Apple chip.

## The engine under most of our numbers {#the-engine-under-most-of-our-numbers}

On 2026-11-12 we read the markdown copy published beside each of the 108 pages on this shelf that carry one: 34 of them name llama.cpp. The doorman at [the print lab](https://research.strata2signal.com/how-the-print-lab-works/) and at [RuleSage](https://rulesage.strata2signal.com/)'s rules desk runs through Ollama, and Ollama runs llama.cpp's own server underneath; [A short history of Mistral](https://research.strata2signal.com/a-short-history-of-mistral/) credited every reading as "taken through Ollama (MIT) over llama.cpp (MIT)". Our two pages on the Strata engine benched llama.cpp itself on three cards. This page is the history behind those numbers.

## One evening in March 2023 {#one-evening-in-march-2023}

On 2023-02-24 Meta announced LLaMA, "a state-of-the-art foundational large language model", in four sizes from 7 to 65 billion parameters, "under a noncommercial license focused on research use cases", with access "granted on a case-by-case basis" to researchers.

Two weeks later, on 2023-03-10, Georgi Gerganov made a commit titled "Initial release" (18:40:58Z) and put it on GitHub as `llama.cpp`. The README he wrote that evening is still in the repository's history: "Inference of Facebook's LLaMA model in pure C/C++"; "The main goal is to run the model using 4-bit quantization on a MacBook"; and, under its list of features, "This was hacked in an evening - I have no idea if it works correctly." It told users to "obtain the original LLaMA model weights" themselves; the project shipped none. How those files got out of Meta's hands that same week, and the licences that followed, is [The weights that got out](https://research.strata2signal.com/the-llama-weights-that-got-out/); this page stays with the program.

That same evening, at 20:01Z, it was on Hacker News, where "Llama.cpp: Port of Facebook's LLaMA model in C/C++, with Apple Silicon support" drew 989 points and 284 comments. The first outside fix was merged that same evening (pull request #3, 23:04Z). From there the record is the repository's own: on 2026-11-12 GitHub counts 134,490 stars and 25,503 forks, under the MIT licence.

## ggml, the library underneath, and ggml.ai, the company {#ggml-and-ggml-ai}

llama.cpp does not do its own arithmetic. Its README says it is built on top of the ggml library, a "tensor library for machine learning" whose repository Gerganov created on 2022-09-18, a week before whisper.cpp (2022-09-25), the same organization's port of OpenAI's Whisper speech model.

Six days in, on 2023-03-16, Gerganov opened a discussion titled "Inference at the edge": "This project will remain open-source", and "Bloating the software with the ideas of today will make it useless tomorrow".

The company came the same year. ggml.ai's own page: "ggml.ai is a company founded in 2023 by Georgi Gerganov to support the development of ggml. Nat Friedman and Daniel Gross provided the pre-seed funding." The GitHub organization `ggml-org`, "AI inference at the edge", was created on 2023-05-22, and llama.cpp, ggml and whisper.cpp live there now; no source we read dates the move from Gerganov's own account, so this page does not.

On 2026-02-20 Gerganov announced "ggml.ai joins Hugging Face to ensure the long-term progress of Local AI": "The ggml team continues to lead, maintain and support full-time the `ggml` and `llama.cpp` libraries and related open-source projects", and "The project will continue to be 100% open-source and community driven as it is now." Hugging Face's own blog announced it the same day: "We are super happy to announce that GGML, creators of Llama.cpp, are joining HF in order to keep future AI open." The ggml.ai page says it in one line: "The company was acquired by Hugging Face in 2026."

## From GGML files to GGUF {#from-ggml-files-to-gguf}

The first format was whatever the converter wrote: a `ggml-model-f16.bin`, then a `q4_0` copy of it. On 2023-06-05 the project merged "k-quants" (#1684), "a series of 2-6 bit quantization methods, along with quantization mixes": the Q2_K to Q6_K types.

On 2023-07-26 Gerganov opened the GGUF pull request (#2398), merged 2023-08-21: a file that describes itself, with key-value metadata beside the weights, and a breaking change, so every existing model had to be converted again. The three formats before it, the pull request's own words, the spec's later versions and what sits inside one of those files are on [Inside a GGUF file](https://research.strata2signal.com/inside-a-gguf-file/), which dates each step from the memory-mapped weights of 2023-03-30 to the spec's landing in November 2023.

## The backends, dated {#the-backends-dated}

The first release ran on the processor. The graphics cards came in through pull requests, each dated here to its merge in GitHub's record.

| Merged (UTC) | Backend, and the pull request | Opened by |
|---|---|---|
| 2023-04-19 | NVIDIA cuBLAS, the card's matrix-arithmetic library, for the batched arithmetic (#1044, "significantly faster than OpenBLAS", its processor-side counterpart) | `slaren` |
| 2023-04-28 | CLBlast over OpenCL, "vendor-agnostic, it runs on basically any GPU" (#1164) | 0cc4m |
| 2023-06-04 | Metal: "full GPU inference of LLaMA on Apple Silicon" (#1642) | `ggerganov` |
| 2023-06-14 | CUDA "full GPU acceleration, KV cache in VRAM" (#1827), the KV cache being the model's running memory of the conversation, now kept on the card | JohannesGaessler |
| 2023-08-25 | HIP, the "ROCm Port" for AMD cards (#1087, opened 2023-04-20) | SlyEcho |
| 2024-01-28 | SYCL, "for Intel GPUs" (#2690) | abhilash1910 |
| 2024-01-28 | Vulkan (#2059, opened 2023-06-30, meant "to eventually supercede the OpenCL backend") | 0cc4m |

The README's "Supported backends" table at master, read 2026-11-12, has seventeen rows, among them CANN, MUSA, Hexagon, WebGPU, an RPC backend and OpenVINO (marked in progress). The readings this page cites ran on CUDA or on a processor.

## The server {#the-server}

A contributor, FSSRepo, added the "Server example with API Rest" on 2023-05-21 (#1443): "I have been working on adding an API to llama.cpp using cpp-httplib." Jeffrey Morgan, one of Ollama's two founders, later said "Ollama was originally inspired by the 'server' example" ([our Ollama history](https://research.strata2signal.com/a-short-history-of-ollama/#the-engine-underneath)). OpenAI-compatible routes followed on 2023-11-25 (#4198).

Today's `llama-server` README lists OpenAI-compatible chat, responses and embeddings routes, an Anthropic Messages route, reranking, "parallel decoding with multi-user support", multimodal input and a web UI; it listens by default on the loopback address, the one a computer uses to talk to itself, so a server that listens only there cannot be reached from another machine. Three additions carry dates on ggml-org's Hugging Face blog: a "router mode" that loads and unloads several models on demand (2025-12-11), the Anthropic Messages API (2026-01-19), and decision models, which answer by scoring the options they are given, a probability for each, instead of writing text (2026-10-02). The v0.8.0 release of 2026-10-28 "wires realtime audio in and out of the server", in its own note's words, through two new routes. On 2026-05-29 Gerganov announced llama.app, a website with a one-line installer, and a single `llama` binary that packs the user-facing tools behind one command; asked the same day, he wrote that "`llama-server` remains as is" and that "`llama serve` and `llama-server` are the same thing". The site says the app is "By the llama.cpp team and Hugging Face".

## What was built on it {#what-was-built-on-it}

**Ollama**'s first README, at v0.0.1 on 2023-07-08, began "Run large language models with `llama.cpp`"; after engines of its own, it went back to `llama-server` on 2026-05-29 (the pull request: "use llama-server exclusively for GGML models"), and the release we read the code at pins llama.cpp build b11081 ([our Ollama history](https://research.strata2signal.com/a-short-history-of-ollama/#what-have-they-shipped)). **LM Studio**'s front page says "Powered by the LM Studio runtime, with MLX and llama.cpp under the hood." **The Strata engine**'s documentation, where its README now sends readers for credits, says "Built with parts of llama.cpp / ggml (MIT)" (despite the similar name, Strata, github.com/Niko1221/Strata, is an unrelated project; strata2signal has no affiliation with it). **Llama Windows**, a tray app Hugging Face announced on 2026-10-13 under the title "Introducing Llama Windows: llama.cpp in your Windows tray", lives in the `ggml-org` organization's own repository (MIT). On Hugging Face, the `ggml-org` organization holds 242 models, 235 of them tagged GGUF by the Hub (read 2026-11-12).

## How it is numbered {#how-it-is-numbered}

A llama.cpp build is named by a count: the build script takes `git rev-list --count HEAD`, the number of commits on master, so build b11514 (published 2026-10-08) is the 11,514th. Those tags are "created on almost every commit to the `master` branch". In August 2026 the project added a second kind: after trial tags from 2026-08-17 (v0.1.2 a pre-release whose note reads "Semantic versioning is still work in progress"), v0.2.0, on 2026-08-21, "marks the beginning of consistent semantic versioning for `llama.cpp`", with `vX.Y.Z` tags for "stable" versions. As of 2026-11-13 (UTC), the day we last read, the newest stable release was v1.0.0, published that day and tagged on the same commit as build b12359. It is the project's first major version. Gerganov announced it on 2026-10-15 in "dev : preparing for v1.0.0", whose "main goal" is "to remove most of the legacy baggage from the APIs", "essentially a breaking change for downstreams and hence the major bump". That post first placed it "sometime by the end of the month or the beginning of November"; the three minor releases its checklist set before it came first (v0.7.0, v0.8.0 and, on 2026-11-04, v0.9.0), and the two pull requests the post names for the removals, "ggml : prune" (#29801) and "llama : prune" (#30565), were merged on 2026-11-12. On 2026-11-11 Gerganov had written in the same thread that after v1.0.0 "major versions would likely happen much more often compared to the time it took us to get to v1", and, of the date: "I think we can do the `v1.0.0` release on Friday 13 Nov." The release note's overview begins "llama.cpp v1.0.0 removes the deprecated public API"; the rest of what it changed is in that note, on the project's [releases page](https://github.com/ggml-org/llama.cpp/releases). Nothing on this page was measured on v1.0.0.

## Where it stands today {#where-it-stands-today}

A project that began as one person's evening is now kept by a team at Hugging Face, in a repository that published its first major version, v1.0.0, on 2026-11-13, announced as a breaking change for the programs built on it, and whose write access its founder tightened this year: on 2026-03-24, "as a security precaution" during a supply-chain attack on another open-source package, Gerganov disabled the write permission of all existing collaborators; an update to the same post restored the maintainers' write access and raised the approvals each pull request needs from one to two. The licence is still MIT.

## What llama.cpp did on our cards {#what-llama-cpp-did-on-our-cards}

Every figure below is printed first on the page it links, with its row, file and field; this page adds none.

- [Qwen3.8-Flash-Next on three cards](https://research.strata2signal.com/flash-next-on-three-cards/) ran llama.cpp at commit `3cf03257` (2026-09-20) on an NVIDIA GeForce RTX 5090 Laptop GPU (24 GB), an RTX PRO 6000 Blackwell (96 GB) and an EVGA GeForce RTX 3090 Ti FTW3 Ultra (24 GB), one request at a time, without the model's [draft head](https://research.strata2signal.com/draft-heads-and-speculative-decoding/): 24.4, 102.889 and 22.490 tokens a second writing the answer to a 4,038-token prompt, under the conditions that page states. On that 24 GB desktop card it also ran a current build, b11514: 22.657 tokens a second beside the pinned commit's 22.490, within that page's margin, so no gap. [The full bench](https://research.strata2signal.com/flash-next-why-the-first-word-comes-sooner/) has every row.
- [Speeds from RAM alone](https://research.strata2signal.com/speeds-from-ram-alone/) ran that build, b11514, on the same files and the same desktop on 2026-11-03 with every layer on the processor and no graphics card in the path: 6.51 tokens a second at 8 threads, against the 22.657 above with the card helping, under the conditions that page states.
- [EmbeddingGemma 2 on release day](https://research.strata2signal.com/embeddinggemma-2-on-release-day/) ran llama.cpp at commit `4fbc76d` (2026-10-06), "built from source as a CPU-only release build", on a laptop's processor.
- [A short history of Ollama](https://research.strata2signal.com/a-short-history-of-ollama/) read Ollama v0.34.4's code, which pins llama.cpp at b11081.

## What to take with you {#what-to-take-with-you}

- **The thing that made local AI ordinary was small.** One C/C++ program with no dependencies, 4-bit weights and a laptop: the README of 2023-03-10 promised no more than that.
- **GGUF is why one model file can run in Ollama, LM Studio and llama.cpp alike.** It has been llama.cpp's format since 2023-08-21; on Hugging Face the `ggml-org` organization alone held 235 models tagged with it on 2026-11-12.
- **A build number is a commit count.** b11514 is the 11,514th commit on master; pin it when you measure, because the project moves on almost every commit: b12328, on 2026-11-12, was 814 commits on.
- **Since 2026-02-20 the ggml team that leads it has been at Hugging Face, and the code is still MIT.** Gerganov's announcement and Hugging Face's blog carry the same date, and the repository's licence field read MIT on 2026-11-12.

## How to check our work — and see it live {#how-to-check-our-work-and-see-it-live}

- **Ask the rules desk.** [RuleSage](https://rulesage.strata2signal.com/), our board-game rules helper, is free and needs no account; the doorman named in this page's first section reads the questions there, through Ollama over llama.cpp's own server.
- **Read the first evening.** The [README at `73c6ed5e`](https://github.com/ggml-org/llama.cpp/blob/73c6ed5e/README.md) is the one quoted above, and the repository's commit list shows all thirteen commits dated 2023-03-10 (UTC), the outside fix at 23:04Z among them.
- **Count a build yourself.** In a full clone of the repository (not a shallow one), `git rev-list --count b11514` prints `11514`, the number in the build's name.
- **Read the server's defaults.** [`tools/server/README.md`](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md) lists every flag with its default, the loopback host among them.

## The rest of the seminar {#the-rest-of-the-seminar}

- [Inside a GGUF file](https://research.strata2signal.com/inside-a-gguf-file/): the format this page dates to 2023-08-21, opened byte by byte.
- [The weights that got out](https://research.strata2signal.com/the-llama-weights-that-got-out/): the LLaMA files this program was written for, how they spread, and the licences that followed.
- [A short history of Ollama](https://research.strata2signal.com/a-short-history-of-ollama/): the server that began with llama.cpp compiled in and came back to llama.cpp's own server in 2026.
- [A short history of vLLM](https://research.strata2signal.com/a-short-history-of-vllm/): the other model server on this shelf.
- [A short history of Mistral](https://research.strata2signal.com/a-short-history-of-mistral/): the model at our doors, and the first of these histories.
- [Qwen3.8-Flash-Next on an RTX 5090 laptop, an RTX PRO 6000 Blackwell and an RTX 3090 Ti](https://research.strata2signal.com/flash-next-on-three-cards/) and [Why the first word comes sooner](https://research.strata2signal.com/flash-next-why-the-first-word-comes-sooner/): llama.cpp beside the Strata engine on three cards, one request at a time.
- [Draft heads and speculative decoding, in plain words](https://research.strata2signal.com/draft-heads-and-speculative-decoding/): what a draft head is, and why a speed figure should say whether one was on.
- [The compressed photograph](https://research.strata2signal.com/the-compressed-photograph/): what storing a model's weights in fewer bits costs.
- [Reading the answer instead of writing it](https://research.strata2signal.com/reading-the-answer/): decision models, the kind ggml-org's blog announced for llama.cpp on 2026-10-02.

## Who ran this, and thanks {#who-ran-this-and-thanks}

Thanks to **Georgi Gerganov** and **the ggml authors** for **llama.cpp** and **ggml** (both MIT), the engine under most of this shelf's language-model numbers, and for a repository whose first evening is still readable in its history; to the contributors named in the backends table, whose pull requests are the record; to **Hugging Face** for its GGUF documentation and the `ggml-org` blog; to **Meta** for the LLaMA announcement this story starts from; to **Ollama** (MIT), **LM Studio** and **the Strata engine**'s author (MIT) for saying what runs under their programs; and to **GitHub** and **Hacker News** for the dated record. None of them owed us anything. A small human team asked for this page, chose what it would and would not claim, and signed off on it; a fleet of AI agents fetched and read every source listed below and drafted it under that team's rulings.

## Sources {#sources}

Every external address below that existed on 2026-11-12 was fetched that day between 21:19 and 21:35 (UTC), when all the counts but one were taken; most were also fetched on 2026-10-14 between 01:05 and 06:29 (UTC) or on 2026-10-29 between 16:21 and 16:27 (UTC). The repository's releases and tags and the v1.0.0 discussion were read on 2026-11-13 between 16:56 and 17:06 (UTC), two of its releases for the first time. Every date on the page is a GitHub field (a commit's author date, a pull request's `merged_at`, a release's `published_at`), a date printed on the page credited, or a sentence quoted from it; our own figures are read from the released pages linked above.

- GitHub — [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) (repository created 2023-03-10T18:58:00Z; "LLM inference in C/C++"; MIT; 134,490 stars and 25,503 forks on 2026-11-12; homepage llama.app); the root commit [`26c08466`](https://github.com/ggml-org/llama.cpp/commit/26c084662903ddaca19bef982831bfb0856e8257) "Initial release", 2023-03-10T18:40:58Z; the [README at `73c6ed5e`](https://github.com/ggml-org/llama.cpp/blob/73c6ed5e/README.md), the last of that evening (2023-03-10T23:30:47Z); [#3](https://github.com/ggml-org/llama.cpp/pull/3), merged 2023-03-10T23:04:06Z; the [README at master](https://github.com/ggml-org/llama.cpp/blob/master/README.md) (the goal sentence, the ggml sentence, the Supported backends table); [`scripts/build-info.sh`](https://github.com/ggml-org/llama.cpp/blob/master/scripts/build-info.sh) (`git rev-list --count HEAD`); [`tools/server/README.md`](https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md) (the features list, the default host)
- GitHub — the pull requests, each dated by its `merged_at`: [#1044](https://github.com/ggml-org/llama.cpp/pull/1044) cuBLAS 2023-04-19; [#1164](https://github.com/ggml-org/llama.cpp/pull/1164) CLBlast 2023-04-28; [#1443](https://github.com/ggml-org/llama.cpp/pull/1443) server 2023-05-21; [#1642](https://github.com/ggml-org/llama.cpp/pull/1642) Metal 2023-06-04; [#1684](https://github.com/ggml-org/llama.cpp/pull/1684) k-quants 2023-06-05; [#1827](https://github.com/ggml-org/llama.cpp/pull/1827) CUDA full offload 2023-06-14; [#2398](https://github.com/ggml-org/llama.cpp/pull/2398) GGUF, opened 2023-07-26, merged 2023-08-21; [#1087](https://github.com/ggml-org/llama.cpp/pull/1087) ROCm 2023-08-25; [#4198](https://github.com/ggml-org/llama.cpp/pull/4198) "server : OAI API compatibility" 2023-11-25; [#2059](https://github.com/ggml-org/llama.cpp/pull/2059) Vulkan and [#2690](https://github.com/ggml-org/llama.cpp/pull/2690) SYCL, both 2024-01-28
- GitHub — the [releases](https://github.com/ggml-org/llama.cpp/releases): [b11514](https://github.com/ggml-org/llama.cpp/releases/tag/b11514) 2026-10-08T20:14:45Z (commit `de7fa0a3`); [v0.2.0](https://github.com/ggml-org/llama.cpp/releases/tag/v0.2.0) 2026-08-21 (the semantic-versioning sentences); [v0.6.0](https://github.com/ggml-org/llama.cpp/releases/tag/v0.6.0) 2026-10-05; [v0.7.0](https://github.com/ggml-org/llama.cpp/releases/tag/v0.7.0) 2026-10-15T17:13:19Z; [v0.8.0](https://github.com/ggml-org/llama.cpp/releases/tag/v0.8.0) 2026-10-28T12:47:51Z, whose note carries the realtime-audio sentence; [v0.9.0](https://github.com/ggml-org/llama.cpp/releases/tag/v0.9.0) 2026-11-04T13:08:06Z; [v1.0.0](https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0) 2026-11-13T10:20:19Z, the release GitHub marked latest on 2026-11-13, whose note carries the "removes the deprecated public API" sentence and whose tag is on commit `b1b770d9`, the commit of [b12359](https://github.com/ggml-org/llama.cpp/releases/tag/b12359), 845 commits ahead of b11514 by GitHub's compare; [b12328](https://github.com/ggml-org/llama.cpp/releases/tag/b12328) 2026-11-12T21:20:31Z, marked pre-release, 814 commits ahead of b11514 by GitHub's compare; the tags `v0.1.0` and `v0.1.1`, made by the project's own automation on 2026-08-17, have no release page, and [v0.1.2](https://github.com/ggml-org/llama.cpp/releases/tag/v0.1.2) (2026-08-18) is a pre-release; and commit [`3cf03257`](https://github.com/ggml-org/llama.cpp/commit/3cf03257f219afbe7334045ff7c6a06ac68c627d), 2026-09-20
- GitHub — the maintainer's discussions: [Inference at the edge](https://github.com/ggml-org/llama.cpp/discussions/205) 2023-03-16; [ggml.ai joins Hugging Face to ensure the long-term progress of Local AI](https://github.com/ggml-org/llama.cpp/discussions/19759) 2026-02-20; [security : limiting write access to the repository/organization](https://github.com/ggml-org/llama.cpp/discussions/20954) 2026-03-24; [llama.app : website + unified `llama` binary](https://github.com/ggml-org/llama.cpp/discussions/23875) 2026-05-29; [dev : preparing for v1.0.0](https://github.com/ggml-org/llama.cpp/discussions/30566) 2026-10-15, its checklist last edited 2026-11-13T10:23:50Z, when v1.0.0 was ticked, with Gerganov's comment in it of 2026-11-11 (the "Friday 13 Nov" sentence) and his reply under it the same day (the "much more often" sentence), and the two pull requests it names, [#29801](https://github.com/ggml-org/llama.cpp/pull/29801) "ggml : prune" and [#30565](https://github.com/ggml-org/llama.cpp/pull/30565) "llama : prune", merged 2026-11-12T18:58:25Z and 18:59:10Z
- GitHub — [ggml-org/ggml](https://github.com/ggml-org/ggml) (created 2022-09-18; "Tensor library for machine learning"; MIT), [ggml-org/whisper.cpp](https://github.com/ggml-org/whisper.cpp) (created 2022-09-25; "Port of OpenAI's Whisper model in C/C++"; MIT), and the organization [ggml-org](https://github.com/ggml-org) (created 2023-05-22; "AI inference at the edge")
- [ggml.ai](https://ggml.ai/) (the founding, funding and acquisition sentences); [llama.app](https://llama.app/) ("By the llama.cpp team and Hugging Face")
- Hugging Face — [GGML and llama.cpp join HF to ensure the long-term progress of Local AI](https://huggingface.co/blog/ggml-joins-hf), 2026-02-20; [GGUF](https://huggingface.co/docs/hub/en/gguf) (the quantization-type table, which sources Q2_K to Q6_K to #1684); the [ggml-org organization](https://huggingface.co/ggml-org) (242 models, 235 tagged GGUF, on 2026-11-12) and its blog: [Model Management](https://huggingface.co/blog/ggml-org/model-management-in-llamacpp) 2025-12-11, [Anthropic Messages API](https://huggingface.co/blog/ggml-org/anthropic-messages-api-in-llamacpp) 2026-01-19, [Decision Models](https://huggingface.co/blog/ggml-org/decision-models-in-llamacpp) 2026-10-02; [Introducing Llama Windows: llama.cpp in your Windows tray](https://huggingface.co/blog/huggingface/llama-windows), 2026-10-13, and its repository [ggml-org/Llama-Windows](https://github.com/ggml-org/Llama-Windows) (MIT)
- Meta — [Introducing LLaMA: A foundational, 65-billion-parameter large language model](https://ai.meta.com/blog/large-language-model-llama-meta-ai/), 2023-02-24
- Hacker News — [Llama.cpp: Port of Facebook's LLaMA model in C/C++, with Apple Silicon support](https://news.ycombinator.com/item?id=35100086), posted 2023-03-10T20:01:54Z by the site's own API record (989 points, 284 comments on 2026-10-14, 2026-10-29 and 2026-11-12)
- [LM Studio](https://lmstudio.ai/) (the front page's sentence); [Strata](https://github.com/Niko1221/Strata) (its README's "Credits and model licenses" link, read 2026-10-29 and 2026-11-12, and the credits section of its `docs/HOW_IT_WORKS.md` it points to: "Built with parts of llama.cpp / ggml (MIT)")
- This workshop's own released pages, linked above: the two Strata pages, EmbeddingGemma 2 on release day, the Ollama, vLLM and Mistral histories, Inside a GGUF file, the print lab, the compressed photograph, [The weights that got out](https://research.strata2signal.com/the-llama-weights-that-got-out/), [Draft heads and speculative decoding, in plain words](https://research.strata2signal.com/draft-heads-and-speculative-decoding/), Speeds from RAM alone; and the count of 34 pages naming llama.cpp, taken on 2026-11-12 over the 108 `index.md` copies on this shelf

Corrections and later measurements will be added below, each dated (UTC), with a window at both ends where one applies and saying in plain words what it counts.

<!-- derived 2026-11-16 (UTC) by tools/derive_md.py from the pour source.
     source html sha256: 159b0fd64ed15f12d6886593ca9d4ee586c2b5c1eeb81bae63fe4a9f314ace4f
     derivation sha256:  d464b98b266ecb8d20a1b0819fcefd687f788fa872f6379a00e6489d1859e85b
     the {#id} on each heading is the anchor that heading carries on the page. -->
