llama.cpp v1.0.0: The Version Number Finally Means Something

After three and a half years of build numbers like b12359, llama.cpp has shipped v1.0.0 — and the headline isn't a feature, it's a promise. The engine underneath a huge share of local AI now carries a semantic version, which means breaking changes to the public C API arrive only with a major-version bump. That's the kind of boring, grown-up commitment that actually matters, because half the local AI ecosystem compiles against this thing.

The catch, and it's a real one: 1.0.0 is itself a breaking release. If your code still calls any of the 35 deprecated functions being removed — or llama_set_warmup, or the defrag-threshold option — it breaks at compile time. The team is using the milestone to clean house first and promise stability second. Fair trade, but don't upgrade blind.

The takeaways

A contract, not a celebration. The versioning rules are refreshingly specific: MAJOR means a breaking change to the public C API, MINOR means a backward-compatible addition, PATCH means a bug fix with no API change. Going forward, API and ABI incompatibilities get major bumps. That turns the version number into a signal you can actually plan around.
The promise has a scope — read it. The contract covers the public C API, the headers under ./include. It says nothing about runtime behaviour staying identical, and the announcement leaves open whether releases change runtime defaults. Pin to a major version and test upgrades against your own workloads anyway.
Speculative decoding graduates. The NextN speculative-decoding API moves from llama-ext.h into the main llama.h header — it's now part of the stable public surface, not an extension. A quiet promotion that says a lot about where the project thinks inference is heading.
Real capability, not just ceremony. FP8 KV caches with OCP FP8 support, a host-memory cache for mixture-of-experts models, external PLE shards loadable from separate GGUF files, and a hardened llama-server — SSRF, cross-origin, DNS rebinding and MCP tool-cancellation fixes, plus an OpenAI-compatible /v1/audio/speech endpoint. The web UI gets a redesign with chat groups and a live-preview markdown editor.
Model coverage keeps widening. Three new text architectures (Kimi-K3.1-Pro, the dense 65B Ripple-65B-Preview, and xing4_0), drama-3 text-to-speech, and vision support for MageVL-Cowork and Kimi-K3.1. Over 150 architecture variants now supported, across backends from NVIDIA and AMD to Qualcomm, Arm, IBM and Intel.

Why this matters

Local AI runs on infrastructure most users never see, and llama.cpp is the load-bearing piece of it. A stable public interface is what lets downstream projects — the apps, tools and products built on this engine — stop chasing a moving target and start building on something that promises not to move without warning. The cleanup hurts once; the contract pays off for the next three and a half years.