Intel / repo

ollama v0.40.15

Source: github.com
Author: github-actions[bot]Top Viber
Date: Published Nov 3, 2026

Context

The mimo-v2.6 model that Ollama added in version 0.40.15 is not the MiMo-V2.6 Pro or Flash model that Xiaomi released in September with claims of matching frontier models on benchmarks. According to its library page, the name refers to MiMo-V2.6-Distill-Qwen-9B, a nine-billion-parameter model that Xiaomi built by Qwen3.5-9B on data generated by MiMo. It is a smaller model trained on examples a MiMo model produced, not the MiMo model itself. Ollama lists it at 6.6 to 8.7 GB to download, with text and image input, , a thinking mode and a of up to 256K tokens, and one command, ollama run mimo-v2.6, fetches and starts it.

The useful comparison is therefore with its Qwen starting point, not with the flagship. The figures on the library page are Xiaomi's own, and the gains vary widely by test. On SWE Pro, built from real software-repair tasks, the score rises from 32.0 to 44.6, and on Terminal Bench 2.1, which has a model complete tasks in a command-line environment, it rises from 27.0 to 37.1. On SWE Verified it moves only from 60.0 to 61.1. Our earlier coverage of the model card reported the same pattern. Anyone pulling this model gets a small local model whose reported edge over its base is large on some agent tasks and negligible on SWE Verified, as measured by Xiaomi rather than an independent tester.

Key takeaways · AI-distilled

Ollama's AutoTrust decision models read text and return choices, yes or no probabilities, and ratings for the questions you supply, invoked with ollama run plus an --ask question.
AutoTrust comes in 9b, 35b and 27b-vl sizes, with the 35b and 27b-vl able to read images; the release notes say autotrust-jev is supported on Apple silicon in this release.
ollama launch offers to migrate existing kimi-cli setups to Kimi Code CLI, its successor, and Claude Desktop's model picker now shows Ollama's actual model names.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.

Read the source github.com