Ollama v0.40.14-rc0 introduces OpenAI API refinements and MLX optimizations

Ollama has released version 0.40.14-rc0, a pre-release update aimed at refining its local AI runtime capabilities. The update introduces several improvements to how the server handles external integrations, specifically targeting OpenAI-compatible API behaviors and cloud-based web search responses.

Technical adjustments were also made to the MLX backend and the core application logic. These changes include better handling of ChatGPT connection failures and improved data processing for long top-k rows within the MLX framework, ensuring more stable local model execution.

Technical improvements in Ollama v0.40.14-rc0

• OpenAI API Compatibility: Updates ensure that model sampler defaults are no longer overridden and that input content in Responses remains readable.
• MLX Backend Optimization: Implements chunk-based sorting for long top-k rows to improve performance on Apple Silicon via the MLX framework.
• Connectivity and Routing: The application now treats ChatGPT disconnect failures as retryable events and verifies ChatGPT routing prior to configuration.
• Server-side Decoding: Added support for decoding cloud web search responses within the server component.
• CI Stability: Pins buildx to resolve a specific deadlock issue identified in the continuous integration pipeline.

Why it matters

These updates improve the reliability of Ollama when used as a local drop-in replacement for OpenAI-style workflows. The MLX optimizations specifically benefit developers running large language models on macOS hardware, while the improved error handling makes the runtime more resilient for production-like local deployments.