Ollama Preview Adds MiMo V2.6 Distilled Model Support, Handling Thinking and Tool-Call Boundaries

Ollama merged PR #19400 on November 3, adding support for MiMo V2.6’s message format. The related commit has appeared in the v0.40.15-rc0 preview tag. Although the proposal’s title refers broadly to model support, comments in the new renderer and the reference template tests point specifically to MiMo-V2.6-Distill-Qwen-9B. The available evidence does not justify extending that claim to full support for the entire Pro or Flash series. Merged pull request, preview tag

Xiaomi’s model card positions this 9B model as a starting point for research in agentic reinforcement learning. It is based on Qwen3.5-9B and was supervised fine-tuned on data generated by MiMo, covering programming, general agent tasks, visual programming, and cybersecurity. The news here is that the model has been integrated into an inference tool; the model itself was already public. Official model card

The core code change recomposes conversation history, thinking content, and tool calls using the boundaries expected by the model. The new renderer writes tool definitions, preserves the assistant’s thinking blocks, and encodes function names and parameters in a compact XML format. On output, it reuses the qwen3.5 parser. New tests feed the same result using different stream chunk sizes and check whether thinking text can be separated correctly from tool names and parameters. From an engineering perspective, this should help reduce the integration risk that an agent treats tool instructions as ordinary text and therefore fails to execute them. Code changes and tests

The adaptation also deliberately diverges from the original template: when thinking is enabled, Ollama prepends the start marker for a thinking block. A code comment explains that this SFT checkpoint may end its turn before closing that boundary. This means deployment results depend on the prompt template as well as the weights. Renderer implementation

Next, watch for inclusion in a stable release and for how the renderer and parser settings work in custom Modelfiles. Maintainer review has cautioned that creating a model directly may apply the existing Qwen3.5 renderer. Public discussion has also raised concerns about looping output, but provides no measurements sufficient to determine whether the issue has been resolved. Engineering teams should first validate multi-turn tool round trips, thinking toggles, and quantized versions before evaluating the model for long-running agent workloads. Maintainer review