llama.cpp Hits 1.0.0 After 3.5 Years, Tying API Breaks to Major Versions
On this page 6 sections
TL;DR
llama.cpp and its tensor library ggml both shipped v1.0.0 on Nov 13, three and a half years after the first commit. The number is a contract: from now on, a breaking change to the public C API in either project means a major-version bump, so anything you build on top can pin a version without betting the build on a moving target. The release itself prunes 35 deprecated functions from llama.h, adds FP8 KV caches and a CPU-side MoE expert cache, lands six new model architectures, and gives llama-server a security pass covering SSRF and DNS rebinding.
What shipped in the box
Version 1.0.0 is a cleanup release as much as a milestone. The deprecated public API is gone: 35 functions, llama_set_warmup, and the defrag_thold setting are all removed, while the NextN speculative-decoding API got promoted from llama-ext.h into llama.h proper. On the ggml side, the deprecated API and the legacy DIAG_MASK ops are pruned too, and the RPC protocol jumped to v9, which rejects incompatible peers outright.
Model support grew too. Three new text architectures landed: Kimi-K3.1-Pro, which is K3 with its full-attention MLA layers replaced by DSA; Ripple-65B-Preview, a dense 65B architecture that needs no new ops; and xing4_0. Vision support arrived for MageVL-Cowork and Kimi-K3.1, drama-3 joined as a TTS architecture, and the server gained an OpenAI-compatible /v1/audio/speech endpoint plus a redesigned web UI with chat groups and a markdown editor.
The contract is the real news
The 1.0 from ggml-org is mostly a promise to stop surprising you. llama.cpp spent three and a half years without promising anything, which is one way to never break a promise. The announcement post ends that: both projects will now reflect API and ABI incompatibilities with major-version bumps, and the team says it will get stricter about those boundaries as adoption grows. The versioning itself had crept in quietly: the previous tagged releases were v0.9.0 for llama.cpp and v0.29.0 for ggml. Both libraries crossed to 1.0.0 the same morning.
An ABI break is the one to watch if you ship binaries. It works like changing the shape of the wall socket: the wiring still carries current, but every appliance built for the old plug now needs an adapter. For everyone who consumes llama.cpp through Homebrew, Debian, conda-forge, MacPorts, Winget, Docker, or the iOS and macOS XCFramework, the version number finally means something you can put in a pin.
llama-server got a security pass
If you run llama-server where other machines can reach it, this release is worth the update on its own. It fixes SSRF through the CORS proxy and media URLs, repairs --cors-origins handling, protects against DNS rebinding, and stops browsers from reaching child instances directly. The HF_TOKEN is now sent only to huggingface.co hosts.
Two capabilities that used to be silently available are now opt-in: --allow-unauthenticated-tools and --media-allow-private-urls. Safer defaults, with the risky switches behind explicit flags, is the right trade for a program whose main job is listening on a port.
Two features for tight VRAM
First, OCP FP8 KV caches: pass -ctk f8_e4m3 and -ctv f8_e4m3 and each cache entry takes 8 bits instead of the 16 bits of F16, with optional static per-tensor scales. Second, the MoE expert cache grew a configurable CPU share: cache misses can be computed in host memory and copied back asynchronously, so expert computation spills into system RAM when VRAM runs short instead of failing.
Neither feature ships with benchmark numbers in the release notes, so measure on your own hardware before committing a config. On a box with a modest GPU and a wall of system RAM, they are the knobs you want.
Key Takeaways
- llama.cpp and ggml both reached v1.0.0 on Nov 13, 2026, three and a half years after the first commit.
- Breaking changes to the public C API now trigger major-version bumps in both projects, which makes a pinned version a real safety net.
- The drop prunes 35 deprecated functions from llama.h, promotes the NextN API into llama.h, and bumps the RPC protocol to v9, which rejects incompatible peers.
- Six new architectures landed, including Kimi-K3.1-Pro and the dense 65B Ripple-65B-Preview, plus an OpenAI-compatible
/v1/audio/speechendpoint. - llama-server fixes cover SSRF, DNS rebinding, CORS and child-instance access, with
--allow-unauthenticated-toolsand--media-allow-private-urlsmoved behind opt-in flags. - FP8 KV caches and a CPU-share MoE expert cache stretch a fixed VRAM budget further on local hardware.
Sources: llama.cpp v1.0.0 release notes, ggml v1.0.0 release notes, ggml-org announcement, tbreak