llama.cpp v1.0.0 is out: after about three and a half years of development, the team behind the popular C/C++ library for local AI inference published its first major version on 13 November 2026. The underlying tensor library ggml reaches version 1.0.0 at the same time. The most important change is not a feature but a promise: from now on, breaking changes to the public C interface will only come with a major version bump. The release also removes deprecated functions, tightens the defaults of llama-server and adds new models and FP8 caches.
What llama.cpp v1.0.0 means for developers
In the announcement on Hugging Face, Daniel Bevenius, Georgi Gerganov and merve explain the move with the growing number of applications built on the library. Every new integration means more developers depending on its interfaces, they write, so a flexible but stable API is essential.
All ggml-org projects now follow semantic versioning rules:
- Major: breaking change to the public C API in
./include - Minor: backward-compatible features, backend updates, new model support or API additions
- Patch: bug fixes with no API change
This applies to the two libraries libllama (headers in llama.cpp/include/) and libggml (headers in ggml/include/). For projects that embed llama.cpp, dependency management becomes much easier: an update within the same major version should not require code changes. Until now, the project was mainly distributed as sequentially numbered builds, sometimes several per day. Those builds continue, and the release also points to nightly build b12359.
Cleanup: these API changes break old code
Before the stability promise takes effect, the team did a thorough cleanup. According to the release notes on GitHub, llama.h drops:
- the function
llama_set_warmup - 35 functions marked as deprecated
- the
defrag_tholdfield in the context parameters and the--defrag-tholdoption
At the same time, the NextN and speculative decoding API moves from llama-ext.h into the main header llama.h. According to the related pull request, that covers 18 functions. llama_set_embd_state_src is renamed llama_set_ctx_state_src. llama_context_params gains the fields moe_cache_size and moe_cache_cpu_share.
Anyone maintaining bindings or linking directly against libllama should check for remaining deprecated calls before switching. The compiler will flag them reliably, because the functions are simply gone. ggml 1.0.0 also removes deprecated public interfaces and the legacy DIAG_MASK operations.
New features: FP8 cache and MoE expert cache
The release brings several additions that mainly help when memory is tight:
- FP8 KV caches: with
-ctk/-ctv f8_e4m3, the key-value cache can be stored in OCP FP8 format, optionally with static per-tensor scales. - MoE expert cache: mixture-of-experts models get a host-memory cache with a configurable CPU share for cache misses and new replacement strategies.
- External PLE shards: additional GGUF shard files can be loaded with
--shardor--ple. - Windows: unbuffered model loading, mirroring the DirectIO path on Linux.
New supported architectures include the text models Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0, the drama-3 text-to-speech model and the vision models MageVL-Cowork and Kimi-K3.1 Vision. According to Hugging Face, llama.cpp now supports more than 150 architecture variants. ggml 1.0.0 also delivers performance work across several backends, including CUDA, Metal, Vulkan, SYCL, OpenCL and WebGPU.
In more detail, the ggml 1.0.0 release notes list configurable YaRN truncation, FP8 support, faster few-row quantized matrix multiplications and MoE fusion on CUDA and Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL and new WebGPU kernels, among other things. How much this changes speed in a given setup depends on the model and hardware. The announcement does not give concrete benchmark figures.
llama-server: stricter defaults
For anyone running llama-server as a local API server, the server changes matter most. Some previously allowed configurations no longer work without an extra flag:
- Enabling tools or MCP while binding the server with
--hostto a non-loopback address without an API key now fails with an error. This has to be allowed explicitly with--allow-unauthenticated-tools. - Media URLs in chat requests can no longer reach local or private addresses by default. Passing images from a NAS on the local network, for example, now requires
--media-allow-private-urls. - The CORS proxy only reaches public addresses. It can no longer be used to reach an MCP server on localhost.
--cors-originsis now handled correctly, including port validation, and the server protects against DNS rebinding.- A configured
HF_TOKENis only sent to huggingface.co hosts.
In the related pull request, the maintainers explicitly say these points are not vulnerabilities. Because many instances now run directly on workstations rather than in a sandbox, they accept a slightly worse user experience. Anyone using scripts or container setups with fixed launch parameters should review them before updating.
The release also adds an OpenAI-compatible /v1/audio/speech endpoint for speech output, cancels MCP tool calls when a client disconnects, and ships a redesigned web UI with chat groups and a live-preview markdown editor.
Distribution, testing and assessment
The team plans to expand distribution. According to the announcement, pre-release artifacts are available for Linux, macOS, Windows, Android and iOS, alongside packages via Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for Apple platforms. For quality assurance, the project relies on Hugging Face services such as HF Buckets and HF Jobs plus self-hosted runners with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.
In practice, version 1.0.0 is mainly a signal to developers who build local models into their own products. Teams that pinned a specific build to avoid breakage can now target a major version instead. Local inference has not become faster or better overnight, but it has become more predictable. That fits the trend toward large open models running on owned hardware, such as Meta Muse Spark Open on two Mac Studios or the open robotics model Runway Praxis-1. Coding tools like Pi 1.0 also show how much stable interfaces now matter for open-source AI.
Recommended upgrade path: build in a test environment first, fix compiler errors caused by removed functions, compare llama-server launch parameters with the new opt-in flags, and only then update production.




