builders · review · first seen Nov 6, 10:33 UTC

Ollama adds MiMo V2.6 and AutoTrust support

Ollama v0.40.15 shipped support for mimo-v2.6, AutoTrust decision models, and concurrent requests for recurrent and vision GGUF models. Builders should care if a local workflow needs a code-oriented model or a structured decision output, but the release notes do not establish either model's quality.

L0 · primary1 source6 claims checked4 min read

What shipped

The release adds support for Xiaomi MiMo's mimo-v2.6, described as a 9B model built for code and agent workflows. The release notes give a direct entry point: ollama run mimo-v2.6. This is support in Ollama, not a reported coding benchmark or a claim that the model will complete a particular agent task. For a builder, the concrete change is a model to try through Ollama. The evidence here is the release artifact, not an independent evaluation. [1]

The same release adds AutoTrust decision models. They take supplied questions and return choices, yes or no probabilities, and ratings. The documented example runs ollama run autotrust-jev with a customer message about a duplicate card charge and asks whether it is a billing issue. That example shows the input pattern: supply text and ask an explicit question. It does not show that a returned probability is calibrated or that the answer should become a final decision without review. [1][2]

AutoTrust comes in 9b, 35b, and 27b-vl sizes. The 35b and 27b-vl sizes can also read images; the family description does not establish image input for the 9b size. The release specifically notes that autotrust-jev is supported on Apple silicon in this release. It also says recurrent and vision GGUF models can serve more than one request at a time. That is a serving-capability change, separate from the new model support. [1][2][3][4]

Why it matters if you build

For a code workflow, the practical test is whether mimo-v2.6 fits the tasks and review standards you already use. Its 9B size and stated purpose make it a candidate to try, but neither predicts correctness on your codebase. Start with a representative task, inspect the result, and keep your existing acceptance checks. The release tells you how to start the model; it does not supply evidence for replacing a current model or removing review from an agent workflow. [1]

For a text decision workflow, AutoTrust offers an explicit question over supplied material. A builder handling incoming requests could ask a narrow classification question, collect the returned choice, probability, and rating, and compare those outputs with decisions already known to be correct. The duplicate-charge example is a useful input pattern, not a validated billing classifier. No threshold, calibration method, error rate, or policy for acting on the outputs is supplied here. Establish those in your own test before an automatic action follows. [1]

Deployment choices are constrained by what the notes actually say. If the question needs an image, look at the 35b or 27b-vl AutoTrust sizes rather than assuming the 9b size reads images. If the machine is Apple silicon, autotrust-jev has explicit support in this release; the note does not settle support on other hardware. Concurrent service for recurrent and vision GGUF models may matter when several requests arrive together, but these claims give no throughput, latency, memory, or operating-cost figures. Measure those on the workload you intend to serve. [1][2][3][4]

How to try it today

1. Run ollama run mimo-v2.6 and give it a representative code or agent task.
2. Review the result against the checks you already apply to that task; do not infer quality from model size.
3. On Apple silicon, run ollama run autotrust-jev with supplied text and an --ask question, as in the release example.
4. Record the choice, yes or no probability, and rating; compare them with an answer you can independently verify.
5. If your application supplies images, evaluate the 35b or 27b-vl AutoTrust size. If it serves recurrent or vision GGUF models, test concurrent requests under your own load.

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. shipped L0 · primary

   Ollama supports mimo-v2.6, a 9B model for code and agent workflows.

   per GitHub releases · ollama/ollama (L0)
2. shipped L0 · primary

   AutoTrust decision models return choices, yes or no probabilities, and ratings for supplied questions.

   per GitHub releases · ollama/ollama (L0)
3. shipped L0 · primary

   The AutoTrust family includes 9b, 35b, and 27b-vl sizes.

   per GitHub releases · ollama/ollama (L0)
4. shipped L0 · primary

   The 35b and 27b-vl AutoTrust models can read images.

   per GitHub releases · ollama/ollama (L0)
5. shipped L0 · primary

   autotrust-jev is supported on Apple silicon in this release.

   per GitHub releases · ollama/ollama (L0)
6. shipped L0 · primary

   Recurrent and vision GGUF models can serve more than one request at a time.

   per GitHub releases · ollama/ollama (L0)

What is not confirmed

These claims do not establish coding accuracy for mimo-v2.6, decision accuracy or probability calibration for AutoTrust, or the performance of any size on a particular task. They give no hardware requirements, download size, memory use, latency, throughput gain, or price. The explicit Apple silicon note for autotrust-jev is not a statement about support elsewhere. Image input is stated for 35b and 27b-vl, not for 9b; concurrent serving is stated for recurrent and vision GGUF models, not as a measured improvement for every model. Treat those as questions for local validation, not release-note promises. [1][2][3][4][5][6]

Who should act, who should wait

Move now if you can run a bounded comparison: a code task for mimo-v2.6, a labeled decision set for AutoTrust, or a concurrent-request test for a recurrent or vision GGUF service. Keep the current workflow as the baseline and check outputs before any downstream action. Wait before switching a production decision path or committing serving capacity if you need established accuracy, calibrated probabilities, resource requirements, or measured concurrency gains. The release shipped support and a serving capability; it did not provide those validation results. [1][2][3]

Sources

• v0.40.15 — GitHub releases · ollama/ollama L0 · primary