llama.cpp Hits 1.0: The Engine Behind Much of Local AI Promises a Stable API, and Clears Out 36 Old Functions to Get There
After three and a half years, the open-source project that runs language models on laptops and servers has tagged version 1.0.0 alongside its ggml tensor library. Major version numbers will now signal breaking changes, but developers who call the C API have a build to fix first, and llama-server is stricter by default.
Published
Last updated / Fact-checked on

Illustrative photo of source code on a screen. It is not llama.cpp code.
The software that quietly powers a large share of "run it on your own machine" AI has reached 1.0.
On Friday, Nov. 13, the ggml-org team tagged llama.cpp v1.0.0 together with v1.0.0 of ggml, the tensor library underneath it. The GitHub release was published at 10:20 UTC, or 2:20 a.m. Pacific. By Saturday, the release had topped Hacker News, according to developer Piotr Sobolewski, and Homebrew listed 1.0.0 as its stable version, according to Shipworthy.
The headline change is not a feature. It is a promise. "Moving forward, both llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps," the team wrote in its launch post on the Hugging Face blog.
At a glance
| Item | Detail | Source |
|---|---|---|
| Release | llama.cpp v1.0.0 and ggml v1.0.0, Nov. 13, 2026 | GitHub, Hugging Face blog |
| Previous release | v0.9.0 | Sobolewski |
| New rule | Major bump = breaking change to the public C API in ./include |
Hugging Face blog |
| Removed | llama_set_warmup plus 35 other deprecated functions, defrag_thold, --defrag-thold, clip.h |
Release notes, Shipworthy |
| Model coverage | More than 150 architecture variants | Hugging Face blog |
| Server changes | Stricter defaults for tools, private URLs and CORS | Shipworthy, release notes |
| License | MIT | Sobolewski |
Why a version number matters
llama.cpp is a C/C++ library that runs language models, including multimodal ones, across hardware from consumer laptops to enterprise servers, using quantization and optimized backends. The team said that in the three and a half years since the project began, "many downstream applications already depend on our work," and that "every new integration means more developers depending on our library interfaces."
For someone chatting with a model in a terminal, a version number is a formality. For the authors of apps and language bindings built on llama.cpp, it is not. Until now, Sobolewski noted, the previous release was still numbered v0.9.0.
Under the new rules, which apply to every ggml-org project using semantic versioning, a major version goes up for a "breaking change to the public C API," a minor version covers "backward-compatible feature, backend updates, model support, or API addition," and a patch is a "bug fix with no API change."
The numbered nightly builds continue alongside the releases, Shipworthy reported, citing the project's release document. A release tag is only created when llama.cpp's internal copy of ggml matches the released ggml, which lets package managers link against a system-wide ggml.
What breaks
To start clean, 1.0 removes old pieces of the C API. The release notes list the removal of llama_set_warmup and 35 deprecated functions from llama.h, 36 in total, all in one pull request, #30565. Shipworthy's guide maps each to its replacement: llama_load_model_from_file becomes llama_model_load_from_file, for example, and llama_free_model becomes llama_model_free. Only llama_set_warmup has no replacement; its deprecation note told users to do warmup runs themselves.
Other changes can also break builds or scripts. The defrag_thold field is gone from llama_context_params, which changes the struct layout, so code built against older headers must be recompiled, Shipworthy said. The --defrag-thold option is gone. The clip.h header has been folded into mtmd, the multimodal library. And llama_set_embd_state_src is now llama_set_ctx_state_src, as the speculative decoding "NextN" API moved into the public llama.h.
People who use llama.cpp through another app or binding get these changes only when that project moves to 1.0, Shipworthy noted. The launch post names no downstream projects and gives them no timeline.
A stricter server
The release also hardens llama-server, the project's built-in API server. According to Shipworthy's reading of the linked pull requests, tools or MCP enabled on a non-loopback host without an API key are now refused unless users opt in with --allow-unauthenticated-tools. Media URLs that point at private network addresses now need --media-allow-private-urls. The web interface's CORS proxy can only reach public addresses, and --cors-origins is now enforced properly.
The reasoning, Shipworthy said, quoting pull request #32115, is that the server was designed on the assumption that users would sandbox it, but "many deployments are now run directly on local computer." The release notes also say HF_TOKEN is now restricted to huggingface.co hosts and that MCP tool calls are cancelled when a client disconnects or times out.
What's new
The release is not only cleanup. The notes list FP8 KV caches, a host-memory cache for mixture-of-experts weights with a configurable CPU share, new text architectures including Kimi-K3.1-Pro and Ripple, drama-3 text-to-speech, new vision models, an OpenAI-compatible /v1/audio/speech endpoint and a redesigned web UI with chat groups. ggml 1.0.0 adds FP8 support and performance work across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon backends.
The team said its CI now runs on Hugging Face infrastructure, with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel. It also addressed AI-written code directly, saying it will bring language models into its development cycle carefully, while keeping code quality, control of the architecture and an understanding of the implementation as its top priorities.
Notably, the announcement includes no benchmarks, no dated roadmap and no stated support window for the 1.x line, as both Sobolewski and Shipworthy pointed out.
Why it matters
Local AI has grown from a hobby into a layer that desktop apps, developer tools and private enterprise deployments build on. A stable, clearly versioned API makes llama.cpp easier to package and safer to depend on, and the server changes reflect that the software is now often exposed on real networks rather than in sandboxes. The trade-off is a one-time migration for anyone still calling the old function names.
What to watch
Watch how quickly major bindings and desktop apps move to 1.0, whether the team publishes a support policy for 1.x, and when the first 2.0 arrives, since that will be the first real test of the new promise. Also watch whether the stricter server defaults cause friction for home and LAN setups.
How we reported this
DailyMech read the ggml-org launch post on the Hugging Face blog, the main sections of the GitHub v1.0.0 release notes, Shipworthy's upgrade guide and Piotr Sobolewski's explainer in full. Details of the server pull requests come from Shipworthy's reading of them. The release came out on Nov. 13; DailyMech is covering it now because it had not reported on it earlier. DailyMech did not build or test the release and contacted no one for this article.
Published November 14, 2026. Facts checked against the cited sources by DailyMech's AI writer.
Sources
- llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem, ggml-org, Hugging Face blog, Nov. 13, 2026 (fully read).
- llama.cpp v1.0.0 release notes, ggml-org, GitHub, Nov. 13, 2026 (main sections read).
- llama.cpp 1.0: what breaks and how to upgrade, Shaheer Q., Shipworthy, Nov. 14, 2026 (fully read).
- llama.cpp 1.0: a stable API after 3.5 years of development, Piotr Sobolewski, Nov. 14, 2026 (fully read).
DailyMech used AI tools to research and draft this article. Facts were checked against the cited sources by DailyMech's AI writer, not by a human editor.