llama.cpp Hits 1.0 After Three and a Half Years, and From Now On a Major Version Means a Broken API

The ggml-org team tagged llama.cpp v1.0.0 and ggml v1.0.0 on November 13, 2026, removing 35 deprecated functions, adopting strict semantic versioning for the public C API, and shipping FP8 KV caches, a host-memory MoE expert cache, new model support and llama-server security fixes.

A white llama stands beside a finished slate-blue and ochre stone arch as its keystone settles into place on a chalky hillside.
A white llama stands beside a finished slate-blue and ochre stone arch as its keystone settles into place on a chalky hillside.

llama.cpp, the C/C++ inference engine that sits underneath a large share of local AI software, has reached version 1.0.0. The ggml-org team published the release on November 13, 2026, alongside ggml v1.0.0, the tensor library that powers it. The MIT-licensed project, which the release blog says is about three and a half years old, had more than 134,000 GitHub stars on release day.

The number matters less as a feature milestone than as a promise to the developers who build on top of it.

A version number that now carries meaning

In the release post on Hugging Face’s blog, the team says many downstream applications already depend on its library interfaces and that a “flexible but stable API is paramount.” From this release on, both llama.cpp and ggml will signal API and ABI incompatibilities with major-version bumps.

The rules apply to all ggml-org projects that use semantic versioning:

  • Major: a breaking change to the public C API in ./include.
  • Minor: a backward-compatible feature, backend update, model support or API addition.
  • Patch: a bug fix with no API change.

The two libraries covered are libllama, with public headers in llama.cpp/include/, and libggml, with headers in ggml/include/. For anyone shipping an app, a desktop wrapper or a language binding, that gives a clear signal for when an upgrade can be dropped in and when code has to change.

Clearing out the old surface first

Version 1.0.0 is itself a breaking release, which is the point of doing the cleanup now. The release notes list the removal of llama_set_warmup, 35 deprecated functions, and the defrag_thold parameter with its --defrag-thold flag. The NextN API, used for speculative decoding, moves from llama-ext.h into the main llama.h header, and llama_set_embd_state_src is renamed llama_set_ctx_state_src. On the multimodal side, clip.h is removed and its model and subcontext code is consolidated into mtmd.

Downstream projects that still call the removed functions will need updates before moving to 1.0.0.

FP8 caches, a MoE expert cache and new models

The tag also carries a substantial batch of new work:

  • FP8 KV caches, selected with -ctk/-ctv f8_e4m3, with optional static per-tensor scales.
  • A host-memory MoE expert cache, exposed through new moe_cache_size and moe_cache_cpu_share context parameters, with a configurable CPU share for cache misses and new replacement policies.
  • Loading of external PLE and engram data from separate GGUF shards via --shard and --ple.
  • New text architectures for Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0, Fish Audio’s drama-3 TTS, and vision support for MageVL-Cowork and Kimi-K3.1, plus Gemma 4 streaming audio input.
  • An OpenAI-compatible /v1/audio/speech endpoint in llama-server.

The blog says llama.cpp now supports more than 150 architecture variants and all mainstream speculative decoding methods.

Opt-ins and SSRF fixes for llama-server

Several llama-server changes are security-focused. The release fixes --cors-origins handling, DNS rebinding, a CORS proxy SSRF issue and child instance access. It adds explicit opt-ins, --allow-unauthenticated-tools and --media-allow-private-urls, for behaviour that now has to be switched on deliberately. The server restricts HF_TOKEN to huggingface.co hosts, and it cancels MCP tool calls when a client disconnects or times out, notifying the MCP server. The bundled web UI is also redesigned, with named chat groups, a live-preview markdown editor and a mobile layout.

Homebrew, Winget and CI runners from six chipmakers

The team lists distribution channels it wants to expand: pre-release artifacts for Linux, macOS, Windows, Android and iOS; Docker; Debian; Homebrew; conda-forge; MacPorts; Winget; and an XCFramework for Apple platforms.

On development, the post says the team continues to integrate LLMs into its development cycle with care, while keeping a high-quality codebase, control of the architecture and an understanding of implementation details as top priorities. It is scaling CI with Hugging Face Buckets and Jobs and adding self-hosted runners with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.

Under the new rules, any 1.x update should drop in without code changes for apps written against 1.0.0, and the next major version will be the signal that something in the public C API has broken.

Topics
  • #llama.cpp
  • #ggml
  • #Local AI
  • #Open Source
  • #Hugging Face
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.