ggml, the small C library that does the heavy math behind llama.cpp and much of the local-AI software built on it, reached version 1.0.0 on November 13. The project Georgi Gerganov started in September 2022 jumped straight from version 0.29.0 to 1.0, and the release notes list 119 changes since then.
The version number is the headline, but the first item in the notes explains it. ggml 1.0.0 removes the deprecated parts of its public interface and two legacy masking operations, DIAG_MASK_INF and DIAG_MASK_ZERO. In software, a 1.0 tag usually signals that the maintainers are willing to commit to a stable interface, and clearing out old calls is the housekeeping that comes first.
What ggml is, in one paragraph
A tensor library handles the grids of numbers that neural networks are made of: multiplying them, reshaping them, squeezing them into smaller formats. ggml does this in plain C with careful memory control, and it runs the same model code on very different chips through a set of interchangeable “backends.” llama.cpp, launched in March 2023, sits on top and handles the language-model specifics. Popular desktop apps for running chatbots offline, Ollama and LM Studio among them, are built around llama.cpp, Wikipedia notes, so changes in ggml eventually reach people who have never heard of it. The team behind both projects, ggml.ai, joined Hugging Face on February 20, 2026.
Where the 119 changes landed
Counting the changelog by backend shows where the work went:
- Nvidia CUDA: about 23 changes, including faster matrix math for batches of 2 to 24 requests, FP8 number support and fixes for mixture-of-experts models.
- Vulkan: about 17, including a FlashAttention path for Intel graphics and better large-matrix settings for AMD’s RDNA 3 and 4 cards.
- OpenCL: about 12, many aimed at Qualcomm’s Adreno phone and laptop GPUs.
- Intel SYCL: about 11, including a mixture-of-experts scheduler that runs on the graphics card itself.
- Apple Metal and Qualcomm Hexagon: about 8 each. WebGPU, for running models in a browser, got about 7.
Those counts are Badiyaa’s tally of commit titles that name each backend; some commits touch several, and four entries are routine syncs with llama.cpp.
Two changes matter beyond any one chip. The scheduler now skips uploading inputs that a step never reads, which cuts wasted copying between memory pools. And the network protocol used to split a model across several machines moves to version 9, loads faster, fails cleanly when a machine cannot be reached, and refuses to talk to older, incompatible peers.
What it means if you run models at home
For most people, nothing changes today. Apps such as Ollama and LM Studio pull in new ggml code on their own schedules, and the speed-ups will arrive with their next updates. The FP8 work and the few-request matrix kernels are aimed at the setup local users actually have: one person, one graphics card, a handful of requests at a time.
Developers who call ggml directly have homework. Code that still uses the removed functions or the old masking operations will need updating, and anyone running the distributed RPC mode has to upgrade every machine together, since mixed versions will now be rejected.
The release lands in a busy month for AI software. Anthropic shipped Claude Fable 5.2 and Mythos 5.2 on November 11, and Linus Torvalds described Linux 7.4-rc1 as a merge window full of AI-found fixes. llama.cpp itself was at version 0.9.0 as of November 4, with more than 109,000 GitHub stars by May; whether it follows ggml to 1.0 is the next version number to watch.
Further reading
Image: AzureSaturn, CC0 via Wikimedia Commons
This article was produced with AI assistance and reviewed by our Editorial Desk. How we work.
