whisper.cpp

A C/C++ implementation of OpenAI’s Whisper automatic speech recognition model with no third-party dependencies: it builds for macOS, iOS, Android, Linux, Windows, WebAssembly and Raspberry Pi, with backends for Metal, Core ML, Vulkan, ROCm, OpenVINO and NPUs. It was published four days after OpenAI announced Whisper, and moved from its author’s account to the ggml-org organisation on 2025-04-04.

Who built itGeorgi Gerganov wrote ggml, the C tensor library this project exists to serve, and llama.cpp; whisper.cpp is his port of OpenAI’s model, created four days after Whisper was announced in September 2022. In February 2025 he asked for maintainers, saying his focus had moved to llama.cpp and ggml, and the repository moved from his personal account to the ggml-org organisation on 2025-04-04. He is still the account with the most commits on it.

1. 01

   Published four days after Whisper

   OpenAI announced Whisper on 2022-09-21, describing a model trained on 680,000 hours of multilingual and multitask supervised data and releasing the model and inference code the same day; the Hacker News post that day drew 1,705 points and 481 comments. whisper.cpp was created on 2022-09-25, and its first commit — `b0a11594aec5`, message “Initial release”, author `ggerganov` — carries an author date of 18:23:15Z, three minutes before the repository’s own `created_at` of 18:26:37Z. The ggml library the project is built on was created on 2022-09-18, and OpenAI’s own repository on 2022-09-16.

2. 02

   The first version number, 83 days later

   The first GitHub release was v1.0.4 on 2022-12-17, 83 days after the repository was created, and six more releases followed between December 2022 and April 2023: v1.1.0 (2023-01-15), v1.1.1 (2023-01-23), v1.2.0 (2023-02-04), v1.2.1 (2023-02-28), v1.3.0 (2023-04-15) and v1.4.0 (2023-04-30). The largest Hacker News threads came from that period: a browser demo on 2023-01-23 with 199 points, and the v1.4.0 release on 2023-04-30 with 162 points.

3. 03

   Long gaps, and what they contained

   Releases then slowed: v1.4.0 (2023-04-30) to v1.4.3 (2023-11-07) is 191 days, and v1.6.2 (2024-05-27) to v1.7.0 (2024-10-05) is 131 days. The v1.7.0 notes list three changes — fixing a crash at high beam counts, lowering VRAM use and speeding up the encoder. v1.8.0 (2025-09-30) leads with “Flash attention is now enabled by default”, and v1.9.0 (2026-06-17) added support for NVIDIA Parakeet. The repository counts 43 releases in total: 39 semantic versions and four nightly builds tagged `b2365`, `b4938`, `b5127` and `b5130`.

4. 04

   The author asks for a maintainer, 2025-02-04

   Two posts went up on 2025-02-04. Discussion #2788, “whisper.cpp : looking for maintainers”, states that the project “has been getting less attention by me because my focus has shifted towards llama.cpp and ggml”, that it “is overall in a good state in terms of the core implementation and functionality”, and that it “is lacking more and more in terms of CI, documentation, packaging and distribution”. Issue #2786, “repo : transfer to ggml-org”, makes the move conditional: “The most important requirement for this transfer to occur is to have an active maintainer(s) for the whisper.cpp project, ideally someone who is or will become part of ggml-org.”

5. 05

   Transferred to ggml-org on 2025-04-04

   Pull request #3005, `rename : ggerganov -> ggml-org`, was opened at 13:06:53Z and merged at 13:11:52Z on 2025-04-04, changing 15 files; issue #2786 closed one second later. The old address `ggerganov/whisper.cpp` returns a 301 to the same repository id, 541269386, whose creation date is unchanged. The ggml-org organisation already existed, created on 2023-05-22 with the description “AI inference at the edge”, and its profile lists `ggml → whisper.cpp` among the projects that depend on the ggml library.

6. 06

   Where it stands

   At the 2026-09-28 snapshot the repository had 53,971 stars, 6,197 forks, 1,042 contributors and 345 open items, with the last commit on 2026-09-24 and v1.9.4 released 2026-09-11. The commit count is still led by `ggerganov` at 1,416, and in the most recent 100 commits (2026-09-18 to 2026-09-24) he made 10, more than any other account in that window. The README’s “Limitations” section has a single entry — “Inference only” — and the WebAssembly example now lives at `ggml.ai/whisper.cpp`.

• The author’s own assessment, published with the maintainer request: the project “is overall in a good state in terms of the core implementation and functionality” while it “is lacking more and more in terms of CI, documentation, packaging and distribution”.
• His stated goal was for the project “to have a long-term maintainer(s) (who is not me) and to continue to live under the ggml-org Github organization”.