Ollama 0.40.15 brings a 9B coding model from Xiaomi, plus AutoTrust decision models that only answer yes, no, or a score

Ollama shipped 0.40.15 on November 6. The project’s rhythm has become interesting: it is less a shell for running models locally and more a place where the parts a local stack needs keep arriving. This release adds two of them, one for writing code and one for making judgments.

mimo-v2.6: 9B, tuned for code and agent loops

ollama run mimo-v2.6 pulls it. It is Xiaomi MiMo’s model, full name MiMo-V2.6-Distill-Qwen-9B: 9B parameters, 256K context, about 6.6GB of download.

The page lists clients it is meant for: Claude Code, OpenCode, Hermes Agent, OpenClaw. All of them are situations where a model runs hundreds of small tool-calling turns. That framing matters more than the parameter count: it targets the many small requests inside an agent loop, not one long generation.

In practice 9B sits in a useful spot, between “runs on a 16GB machine” and “not too stupid to emit valid JSON”. For real agent work, a local 9B with a cloud model as fallback beats both all-cloud (cost) and all-local (reliability).

AutoTrust: another round of decision models

The second addition is AutoTrust, described briefly: it reads text and returns choices, yes/no probabilities, or ratings for the question you supply.

ollama run autotrust-jev "A customer says their card was charged twice." --ask "Is this a billing issue?"

The family ships 9b, 35b and 27b-vl sizes; the latter two can read images. The d3 models in 0.40.11 went the same way, and now AutoTrust joins them, which suggests vendors treat “small models that classify” as a category of its own.

My take is simple: these models earn their keep through stability, not intelligence. A ticket router that is right 93% of the time with the same reasoning every run is more useful than one that is right 96% with a different story each time, because you can write regression tests, tune thresholds and roll it out gradually. You cannot unit test a model that answers by mood.

The smaller fixes that matter day to day

• ollama launch now supports Kimi Code CLI and offers to migrate an existing kimi-cli setup
• Claude Desktop’s model picker finally shows real Ollama model names instead of internal identifiers
• Recurrent and vision GGUF models can serve more than one request at a time; these used to serialize
• Long-context (YaRN) scaling is fixed for Qwen3 and Qwen3.5
• Running ollama with no arguments opens the welcome launcher

How I would use the two

Job	Tool	Why
Small edits, test scaffolding	mimo-v2.6	9B is fast enough to run on a laptop
“Should CI block this?”	a decision model	Boolean output, testable
Long document summarization	a large cloud model	a local 9B drops detail

The real change in local inference over the last two years is not that models got smarter. It is that they got divided up. A pipeline with a small router, a large generator and a decision model is standard practice in the cloud, while locally we had “one big model and a chat” until recently. Releases like 0.40.15 are the missing parts arriving one at a time.