Open sourceNotable

llama.cpp and ggml reach v1.0.0, and breaking changes now get a major-version bump

After roughly three and a half years of numbered nightly builds, the local-inference project commits to semantic versioning for its C libraries and uses the 1.0 release to clear out its deprecated API.

Promptea Editorial4 min read

The maintainers of llama.cpp published v1.0.0 of the project on Friday, together with v1.0.0 of ggml, the tensor library underneath it. The announcement came as a post on the Hugging Face blog signed by Daniel Bevenius, Georgi Gerganov and merve of ggml-org, alongside a tagged GitHub release marked as the project's latest.

The version number is the news. For most of its roughly three and a half years, llama.cpp shipped as a continuous stream of builds identified by a number (the GitHub releases page still lists nightly builds such as b12370), and downstream projects had no simple signal for when an update would break them. The maintainers say that many applications now depend on the libraries and that v1.0.0 reflects "our responsibility to those developers."

What the version numbers now promise

The blog post sets out a semantic-versioning policy for both libraries, `libllama` (public headers in `llama.cpp/include/`) and `libggml` (headers in `ggml/include/`):

  • Major: a breaking change to the public C API in the include directory.
  • Minor: backward-compatible features, backend updates, new model support or API additions.
  • Patch: bug fixes with no API change.

The prose of the post goes slightly further, saying API and ABI incompatibilities will be reflected in major-version bumps; the rules table itself only names the C API. The post does not describe a release cadence or a long-term-support window, so v1.0.0 is a commitment about how breakage will be labelled, not about how long a given version will be maintained.

A clean-out before the promise starts

Fittingly, the 1.0 release is itself a breaking one. According to the release notes, it removes `llama_set_warmup`, 35 deprecated functions, and the `defrag_thold` parameter along with its `--defrag-thold` command-line flag. It also renames `llama_set_embd_state_src` to `llama_set_ctx_state_src` and promotes the NextN speculative-decoding API from an extension header into `llama.h`. Anyone linking against libllama directly should expect to touch code when moving from the 0.x line; the notes list the removals but do not include a migration guide.

ggml v1.0.0 likewise removes deprecated APIs and the legacy `DIAG_MASK` ops, adds configurable YaRN truncation and FP8 support, and bundles performance work across the CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon backends.

What else ships in the release

  • FP8 KV caches, selectable with `-ctk/-ctv f8_e4m3`, with optional static per-tensor scales. An 8-bit cache takes half the memory of the default 16-bit one, which matters most for long contexts; the notes do not publish quality measurements.
  • A host-memory MoE expert cache, configured through new `moe_cache_size` and `moe_cache_cpu_share` context parameters, aimed at running mixture-of-experts models whose weights do not fit in GPU memory.
  • New model support, including Kimi-K3.1-Pro, Ripple-65B-Preview, xing4_0 and the drama-3 text-to-speech model, plus vision support for Kimi-K3.1 and MageVL-Cowork.
  • llama-server changes: fixes for CORS, DNS rebinding and a CORS-proxy SSRF issue, explicit opt-ins for unauthenticated tools and private media URLs, an `HF_TOKEN` restricted to huggingface.co, a new `/v1/audio/speech` endpoint, and a redesigned web UI.
  • Structured-output plumbing: `response_format` support in more chat parsers, JSON-schema and tool-argument fixes, and conversion of Responses-style `text.format` schemas.

The maintainers also say they are scaling CI with Hugging Face Buckets and Jobs, with self-hosted runners backed by hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel, and that llama.cpp covers more than 150 architecture variants. Those figures come from the project itself.

What it means for developers

For people who run models locally through tools built on llama.cpp, little changes overnight; those tools will pick up the release on their own schedule. The practical shift is for anyone who embeds `libllama` or `libggml` directly or packages them. Until now, pinning a build number was the only defence against silent breakage. From here, the major version is meant to carry that signal, which makes ordinary dependency ranges usable.

The server changes are worth a look on their own. If you expose `llama-server` beyond localhost, or rely on it for JSON-schema-constrained output through an OpenAI-compatible client, test the upgrade: several of the security fixes change defaults to opt-in, and the structured-output fixes may change how existing schemas are parsed.

Independent reporting on the release could not be opened from Promptea's research environment at publication time, so the details above rest on the project's own blog post and release notes.

Why this matters

  • llama.cpp sits underneath a large share of local-inference tools, so a versioning promise from it affects how much of that ecosystem manages upgrades.
  • The 1.0 release removes deprecated functions and renames others, so code linking libllama directly will likely need changes.
  • FP8 KV caches and the MoE expert cache target the two common local bottlenecks: long-context memory and large mixture-of-experts models.

Key takeaways

  • llama.cpp v1.0.0 and ggml v1.0.0 were released on November 13, 2026.
  • Breaking changes to the public C API will now require a major-version bump; nightly numbered builds continue alongside tagged releases.
  • The release removes llama_set_warmup, 35 deprecated functions and defrag_thold, and adds FP8 KV caches, a MoE expert cache and llama-server security fixes.

Sources

  1. Hugging Face (ggml-org)Primary
    llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem
    huggingface.co
  2. GitHub (ggml-org/llama.cpp)Primary
    Release v1.0.0 · ggml-org/llama.cpp
    github.com

How we choose and verify sources

Tags:
  • llama.cpp
  • ggml
  • local inference
  • semantic versioning
  • open source
  • KV cache
Companies:
  • ggml-org
  • Hugging Face
Models:
  • Kimi-K3.1-Pro
  • Ripple-65B-Preview
llama.cpp v1.0.0: semantic versioning and breaking API cleanup · Promptea