llama.cpp 1.0: a stable API after 3.5 years of development

On November 13, 2026, the ggml-org team released llama.cpp 1.0 (v1.0.0, to be exact) together with ggml v1.0.0. The announcement on the Hugging Face blog is signed by Daniel Bevenius and Georgi Gerganov, among others, and a day later the ggml release sat at the top of Hacker News.

What changes in llama.cpp 1.0?

The biggest change is a promise rather than a feature. From this version on, both projects follow semantic versioning, and breaking changes to the API and ABI are supposed to be signalled by a bump of the major version number. There are three rules, and they apply to every ggml-org project:

1. MAJOR goes up when a change breaks the public C API (the headers in the include directory).
2. MINOR covers backward-compatible additions: features, backend updates, support for new models.
3. PATCH is for bug fixes with no API changes.

If you only chat with a model in a terminal, that sounds like a formality. For authors of apps and bindings built on llama.cpp it matters a lot: the previous release was still numbered v0.9.0. According to the announcement, 3.5 years after the project started, so much software depends on it that a flexible yet stable API has become a first-order concern.

The cleanup has a price. According to the Release Notes, 35 deprecated functions are gone from llama.h, llama_set_warmup and the --defrag-thold option were removed, llama_set_embd_state_src is now called llama_set_ctx_state_src, and the clip.h header has been absorbed into mtmd. ggml lost its deprecated public API and the old DIAG_MASK operations. Anyone with their own code on those functions, or scripts that pass the removed flag, has to fix them before upgrading.

What else is in the release?

This is more than a housekeeping release. The Release Notes also list:

• support for new models: Kimi-K3.1-Pro, Ripple-65B-Preview, xing4_0 and the drama-3 TTS model from Fish Audio,
• a new server endpoint, /v1/audio/speech, compatible with the OpenAI API,
• an FP8 KV cache (-ctk/-ctv f8_e4m3) and an extended cache for MoE (mixture-of-experts) models that can compute some of the misses on the CPU,
• security fixes in the server: --cors-origins, DNS rebinding and SSRF through the CORS proxy, plus HF_TOKEN now being sent only to huggingface.co hosts,
• performance work across the ggml backends, from CUDA and Metal to Vulkan, SYCL, OpenCL, WebGPU and Hexagon.

According to the announcement, llama.cpp now supports more than 150 architecture variants and all the popular speculative decoding methods. Quality is meant to be guarded by an extensive CI setup: every change landing on master is checked on Hugging Face infrastructure, and with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel the team keeps adding its own test machines.

How do you try llama.cpp 1.0?

The project README gives a one-line install. On macOS and Linux:

curl -LsSf https://llama.app/install.sh | sh

In PowerShell on Windows:

irm https://llama.app/install.ps1 | iex

After that, you point it at a model on Hugging Face. The README example downloads a small Qwen3.5 and opens a chat or a local server:

llama cli -hf ggml-org/Qwen3.5-0.8B-GGUF
llama serve -hf ggml-org/Qwen3.5-0.8B-GGUF

For people who prefer windows to terminals there is the free Llama app for Mac and Windows 11, developed by the llama.cpp team and Hugging Face. It has a built-in chat and a local API at localhost:9931/v1 that works with the OpenAI SDKs. The app’s site also gives memory recommendations: 16 GB for Gemma 4, 24 GB for GPT-OSS and 32 GB for Qwen 3.8 (check how much free memory you have before downloading a bigger model).

The project is also in Homebrew, Winget, conda-forge and MacPorts, and ships as Docker images and Debian packages. If local models are new to you, I described a gentler start in the guide on how to run AI locally.

What should you watch out for?

First, what the announcement leaves out. It has no benchmarks and no dated roadmap of upcoming features. Version 1.0 speaks to the project’s maturity, and it doesn’t claim a jump in performance.

Second, the scope of the promise. Breaking changes will still happen; they are just supposed to be marked by the version number. That covers the public C API, and it doesn’t extend to every flag in the tools.

Third, the approach to writing code with AI. The team says it will bring language models into its work in a deliberate way, with code quality and control over the architecture staying the priority. In a project that so many other people’s apps stand on, that’s probably the right order ;)

The Hacker News comments are mostly congratulations. One user called ggml and llama.cpp “the Linux of the AI world”, and another wrote that on an AMD card that supports only Vulkan, their software is usable thanks to ggml and llama.cpp.

FAQ

Is llama.cpp 1.0 free?

Yes. It is an open-source project under the MIT license, and the Llama app is free too. You pay only in electricity (and wear on your hardware).

Will anything stop working after upgrading to llama.cpp 1.0?

It may, if you use the removed pieces: the 35 deprecated functions from llama.h, the clip.h header or the --defrag-thold option. The full list is in the Release Notes.

What is the difference between llama.cpp and ggml?

ggml is a tensor library for machine learning, the layer that runs the computation on different hardware. llama.cpp is a C/C++ library built on top of it for running language models, with tools such as a terminal chat and a server.