Ollama v0.40.15 adds MiMo V2.6 support and concurrent GGUF request handling

Ollama has released version 0.40.15, introducing support for several new specialized model families. This includes Xiaomi's MiMo V2.6, a 9B parameter model optimized for coding and agentic workflows, and the AutoTrust decision model family, which provides probabilistic ratings and yes/no responses for text and image inputs.

Beyond model additions, the update improves the runtime's efficiency and integration capabilities. Key technical improvements include the ability for recurrent and vision-based GGUF models to handle multiple simultaneous requests and support for long-context YaRN scaling for Qwen3 and Qwen3.5 models. The release also enhances the user experience for CLI and desktop integrations like Kimi Code CLI and Claude Desktop.

Technical improvements in Ollama v0.40.15

• Model Support: Added Xiaomi MiMo V2.6 (9B) for code/agent tasks and AutoTrust decision models (9b, 35b, and 27b-vl), with the latter supporting Apple silicon.
• Concurrency Enhancements: Recurrent and vision-based GGUF models are now capable of serving more than one request at a time.
• Context Scaling: Implementation of long-context support using YaRN scaling for Qwen3 and Qwen3.5 model architectures.
• Tooling Integration: Improved interoperability with Kimi Code CLI (including migration paths) and automated model list population for Pi and Claude Desktop.

Why it matters

By enabling concurrent requests for vision and recurrent GGUF models, Ollama increases its utility for developers building multi-user or high-throughput local AI applications. The addition of specialized decision models and long-context scaling further positions the runtime as a robust platform for complex, agentic local AI workflows.