npx @buildinternet/releases get ollamaRecent releases shift Apple Silicon to MLX inference by default and introduce a structured decision-model API for local classification and routing tasks.
Apple Silicon now defaults to MLX inference. Supported model architectures on Apple Silicon devices run on the MLX backend automatically — no flags or environment variables required. Decision models (Nimble, Tev1, Clef, Clef Flash) and the embeddinggemma-2 embedding model are available on MLX as well, expanding the range of workloads that can use the optimized path.
Decision models return structured answers instead of generated text.
The /v1/systemone endpoint accepts a context state and one or more typed questions, returning choices, probabilities, and scores. Use cases include ticket triage, model routing, and content classification. Cloudflare's Clef (27B) and Clef Flash (9B) models are multimodal — requests can include images alongside the text state, scored jointly with it — and both are available through the same endpoint.
Web search depth per response increased. Models using web search can now perform up to ten searches per response, up from a previous limit of three.
Background model upgrade overhead is paused. Background model upgrade processing, which added per-request overhead in high-volume scenarios such as embedding workloads, is paused in v0.40.3 while that cost is evaluated.
Cloud usage and balance APIs are now proxied. The server now proxies Ollama cloud usage and balance APIs, letting the local server relay account status from hosted services without a separate client call.
OpenAI-compatible Responses history now accepts custom tool calls, and the MLX Metal custom kernels are ported to CUDA. Cloud payment errors are clarified in the app, plus docs standardization and an MLX version bump.
*…