llama.cpp reaches v1.0.0 with a stable API promise, FP8 caches and a hardened local server
The ggml-org team released llama.cpp v1.0.0 and ggml v1.0.0 on November 13, 2026, committing to semantic versioning so breaking changes to the public C API only arrive with a major-version bump, alongside new model support, FP8 KV caches and security fixes for llama-server.
What happened
- llama.cpp and its ggml tensor library both moved to v1.0.0; from now on, breaking changes to the public C API mean a major-version bump, new features and model support a minor bump, and bug fixes a patch bump.
- The release removes the deprecated public API (35 deprecated functions in llama.h) and promotes the speculative-decoding (NextN) API into llama.h, so apps built on older interfaces may need code changes.
- New features include FP8 KV caches, a host-memory cache for mixture-of-experts layers with a configurable CPU share, and support for Kimi-K3.1-Pro, Ripple-65B-Preview, xing4_0, drama-3 text-to-speech and Kimi-K3.1 vision.
- llama-server gets security fixes for CORS, DNS rebinding and a CORS-proxy SSRF issue, new opt-in flags for unauthenticated tools and private media URLs, an OpenAI-compatible /v1/audio/speech endpoint, and cancellation of MCP tool calls when a client disconnects.
- The built-in web UI was redesigned with named chat groups, a live-preview markdown editor and a mobile layout; the Hugging Face blog post says the project now supports more than 150 model architecture variants.
Why it matters: llama.cpp sits underneath many of the apps people use to run AI models privately on their own computers, so a stable, versioned API should mean fewer surprise breakages in those apps. If you run llama-server on a network, the security fixes are a reason to update.
Sources
- GitHub: llama.cpp v1.0.0 release notes github.com
- Hugging Face blog: llama.cpp v1.0.0 huggingface.co
Checked November 15, 2026 against the sources above. Claims about performance or benchmarks are the company's own unless stated otherwise.