llama.cpp 1.0.0 is an inference engine for running large models locally — not a chat app, and not a model itself. Released on November 13, 2026, version 1.0.0's biggest change is a versioning promise: breaking changes to the public C API will bump the major version, compatible new features the minor version, and bug fixes the patch version. Beginners don't need to rush into compiling it — first figure out which path fits you, then decide whether to upgrade.
Three Things to Get Straight First
- It only "runs" models: to chat, pair it with llama-server, llama.app, or a higher-level tool like Ollama or LM Studio. Downloading the engine alone won't get you chatting.
- Models must be in GGUF format: llama.cpp runs GGUF model files, so not just any model weights will work. Check the format before downloading.
- 1.0.0 is a "stability promise", not "faster": the release makes no speed claims, so don't upgrade for performance. Check the removals below first to see whether they affect you.
Check These Two Removals Before Upgrading
1.0.0 removes 35 deprecated public functions and llama_set_warmup, along with defrag_thold and the --defrag-thold startup flag. If you reuse old launch scripts or have written code against the API, search them for these names first — any hit must be fixed, or the upgrade will fail outright. The NextN speculative decoding API moves from llama-ext.h into the main header llama.h, so developers should update their includes accordingly. On the new-feature side, 1.0.0 adds OCP FP8 KV cache, host-memory MoE expert caching, and external PLE shard loading, plus support for new architectures such as Kimi-K3.1-Pro. llama-server gains an OpenAI-compatible /v1/audio/speech endpoint, and fixes security issues including CORS and DNS rebinding.
Three Ways for Beginners to Get Started
- Easiest: use the prebuilt binaries or llama.app. The 1.0.0 prebuilds cover Linux, macOS, Windows, Android, and iOS, with Docker, Homebrew, conda-forge, and other channels — no compiling needed.
- To run it as a service: start llama-server, which exposes an OpenAI-compatible interface you can connect to a chat front end.
- To modify it yourself: pull the source from the ggml-org llama.cpp repository on GitHub and build it. That path suits developers; ordinary users don't need it.
1.0.0 has just shipped, and old bindings and scripts need time to catch up, so don't rush the upgrade in production — get it working on a test machine first. The web UI has changed a fair amount too: the sidebar can fold away and chats can be grouped. A new interface doesn't mean everything works differently, so look for things where you expect them first, and only then explore the new layout.
