Ollama v0.30.0リリース――llama.cpp直接サポートとGGUF互換性を実現 v0.30.0-rc32: llama-server followups (#16353)

Ollama has shipped v0.30.0, and while the update may look incremental from the outside, the internal architecture changes are substantial. The most significant shift is a move away from building on top of the GGML tensor library toward directly integrating llama.cpp as the inference backend. For end users the interface remains familiar, but the change sets the stage for faster adoption of upstream optimizations and new model support.

llama.cpp, created by Georgi Gerganov, has become the de facto open-source engine for running large language models locally. Originally tightly coupled with GGML — its underlying tensor math library — llama.cpp has evolved into a broader project with its own optimizations, hardware backends, and model support that now extends well beyond what GGML alone provides. By integrating llama.cpp directly, Ollama positions itself to track upstream improvements more closely rather than relying on an intermediate abstraction layer.

Alongside this architectural change, v0.30.0 brings native compatibility with the GGUF file format. Introduced in 2023 as a successor to the older GGML format, GGUF stores model weights alongside richer metadata including quantization details, tokenizer configuration, and architecture parameters in a single self-contained file. Hugging Face and other model repositories have largely standardized on GGUF for community model distribution, so improved support here means users should find it easier to pull in third-party models without conversion headaches.

For macOS users on Apple Silicon, the release introduces MLX-based acceleration. MLX is Apple's open-source machine learning framework designed to exploit the unified memory architecture of M-series chips, where CPU and GPU share the same memory pool. This avoids the data-copy overhead typical of discrete GPU setups and can meaningfully improve throughput for local inference workloads. Apple has been quietly building out the MLX ecosystem, and its adoption in Ollama adds another practical showcase for the framework beyond research use cases.

Ollama competes in an increasingly crowded space of local LLM runners that includes LM Studio, Jan, and llama.cpp itself used directly via CLI. Ollama's appeal has always been its low-friction setup — a single binary, a simple REST API, and a model registry that abstracts download and configuration. The architectural improvements in v0.30.0 reinforce that positioning by ensuring the project stays current with the fast-moving llama.cpp ecosystem rather than lagging behind it.

It is worth noting that this release carries the designation rc32, suggesting it is still formally a release candidate and that final stability validation may be ongoing. Users in production environments may want to confirm stability before upgrading, though the high rc number implies the codebase has undergone extensive testing cycles.