Dev & code

llama.cpp Hits 1.0 After Three and a Half Years: A Stable API, 35 Removed Functions, and a Hardened Server

The local inference engine llama.cpp and its computational library ggml reached version 1.0.0 on November 13. The main shift is a commitment: any breaking change to the public C API will now bump the major version number. The immediate tradeoff: 35 deprecated functions have been removed, requiring downstream applications built on the library to make code adjustments.

A streamlined clockwork mechanism with redundant gears removed, leaving a robust and polished central core.
AI-generated illustration

If you run a language model locally on your machine, chances are llama.cpp is doing the heavy lifting under the hood: it is the C/C++ inference engine powering many local AI applications. After three and a half years in “0.x” territory, the project and its underlying compute library ggml reached version 1.0.0 on November 13, announced on the Hugging Face blog, whose company employs the team.

A Commitment More Than a Feature

The real shift here isn’t technical: it’s a versioning pledge. Moving forward, any breaking change to the public C API (the headers in the include folder) will trigger a major version bump; backwards-compatible additions, such as a new backend or a new model architecture, will bump the minor version; and bug fixes will update the patch number. Until now, as the team pointed out in the planning discussion, breaking changes were routinely pushed in 0.x releases without formal notice. The team itself cautions that major versions will likely land far more frequently than the time it took to reach 1.0: the promise isn’t “no more breaking changes,” but “visible breaking changes.”

The Cost of Housekeeping

To clear technical debt, 1.0 breaks quite a bit in one go. According to the release notes, llama.h drops 35 deprecated functions, llama_set_warmup, and the defrag_thold setting (along with the --defrag-thold CLI flag). llama_set_embd_state_src has been renamed to llama_set_ctx_state_src, and the clip.h header has been removed in favor of mtmd. Over on ggml, the old deprecated API and DIAG_MASK operations were dropped, and the RPC protocol has been bumped to v9, rejecting incompatible peers. Any application or language binding calling these legacy functions, or any script passing the removed CLI flag, will need patching before upgrading.

What Comes With It

The release also brings tangible new features: FP8 KV caches (-ctk/-ctv f8_e4m3) to trim context memory overhead, a host-memory MoE expert cache with partial CPU computation offloading, and support for new models, including Kimi-K3.1-Pro and Ripple-65B-Preview, two open models we previously covered. The built-in llama-server has been hardened against several web-based attack vectors (SSRF, CORS, DNS rebinding), cancels MCP tool calls when the client disconnects, and strictly restricts the Hugging Face token to Hugging Face hosts. It also gains an OpenAI-compatible speech synthesis endpoint (/v1/audio/speech) and a redesigned web interface. The project boasts support for over 150 architecture variants, with packages distributed via Homebrew, Winget, Debian, conda-forge, MacPorts, and Docker.

Our take

For anyone running llama.cpp via a packaged end-user application, 1.0 changes almost nothing day-to-day, and the announcement promises no performance gains. However, it is a significant milestone for developers building on top of it: a trustworthy major version finally lets teams lock in dependencies (“1.x”) without worrying that a minor update will break their build, bringing llama.cpp closer to standard system library practices. The hardening of llama-server is the second major takeaway: this server often runs on developer workstations hooked up to MCP tools—a setup notoriously vulnerable to malicious websites targeting localhost. Practical tip: pin version 0.9 until your upstream dependencies migrate to 1.0, and double-check the list of removed functions before upgrading.

Sources

  1. Hugging Face (ggml-org blog) — llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem (November 13, 2026) Primary source
  2. GitHub — ggml-org/llama.cpp, release notes v1.0.0 Primary source
  3. GitHub — ggml-org/ggml, release notes v1.0.0 Primary source
  4. GitHub — llama.cpp, discussion #30566 “preparing for v1.0.0” Primary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.