Ollama 0.40.15, released November 3, 2026, adds Xiaomi’s 9B MiMo V2.6 coding model and AutoTrust decision models, brings Kimi Code CLI to ollama launch, and lets recurrent and vision GGUF models serve concurrent requests. Version 0.40.14, a day earlier, fixed Claude Code output limits and sampler defaults.
Last updated: November 8, 2026
Table of contents
- What is new in 0.40.14 and 0.40.15
- MiMo V2.6 for local coding agents
- AutoTrust decision models
- How to update and try the new features
- ollama launch changes
- Limitations and upgrade notes
- Frequently asked questions
- Sources
What is new in Ollama 0.40.14 and 0.40.15
Ollama shipped two releases within about 21 hours. According to the 0.40.14 release notes, published November 3, 2026 at 01:40 UTC, that version focused on fixes for coding agents and API compatibility. The 0.40.15 release notes, published at 22:52 UTC the same day, added two model families and several integration changes. If you are coming from October, our guide to Ollama’s October 2026 features covers speech models and earlier decision models.
| Change | Version | Who it affects |
|---|---|---|
| MiMo V2.6 (9B, code and agents) supported | 0.40.15 | Local coding agent users |
| AutoTrust decision models (9B, 35B, 27B-VL) | 0.40.15 | Routing, classification and grading pipelines |
| Kimi Code CLI in ollama launch, with kimi-cli migration | 0.40.15 | Kimi users |
| Recurrent and vision GGUF models serve more than one request at a time | 0.40.15 | Server and multi-user setups |
| YaRN long-context scaling for Qwen3 and Qwen3.5 | 0.40.15 | Long-document work |
| Running ollama with no arguments opens a welcome launcher | 0.40.15 | New users |
| Claude Code gets the model’s own output token limit | 0.40.14 | Claude Code with cloud models |
| OpenAI-compatible endpoints keep packaged sampler defaults | 0.40.14 | API clients |
| Faster top-k sampling on MLX for large vocabularies | 0.40.14 | Apple silicon |
MiMo V2.6 for local coding agents
The mimo-v2.6 library page describes MiMo-V2.6-Distill-Qwen-9B as Xiaomi MiMo’s 9B model for code and agent workflows. It was fine-tuned from Qwen3.5-9B on MiMo-generated data, accepts text and images, and supports thinking and tool use. The download is 6.6 GB to 8.7 GB depending on the variant, and the context window is 256K tokens. It is released under the MIT license.
Xiaomi reports gains over the Qwen3.5-9B base, as listed on the library page: SWE Pro rises from 32.0 to 44.6, Terminal Bench 2.1 from 27.0 to 37.1, and SWE Verified from 60.0 to 61.1. These are publisher figures, not independent tests, so measure them on your own repositories.
The page lists one-command setups for ollama launch claude --model mimo-v2.6, OpenCode, Hermes Agent and OpenClaw. A 9B model with a 256K window is a practical choice for a laptop agent that reads long files or screenshots of an interface.
AutoTrust decision models
Decision models answer a question about a piece of text with a probability for every option, instead of free-form prose. The autotrust-jev library page says AutoTrust’s models return choices, yes or no probabilities, and ratings on six levels, and several questions can be asked about the same input in one request. Use them to route support tickets, classify documents, or grade an answer against your own criteria.
The default autotrust-jev tag runs the 9B text model (11 GB). For images, choose autotrust-jev:27b-vl or autotrust-jev:35b (26 GB). The models list a 256K context and use the Apache 2.0 license. AutoTrust reports a yes or no AUROC of 0.996 and a rating mean absolute error of 0.103 on its own held-out test set. Over HTTP the models use the /v1/systemone endpoint, and the Python and JavaScript libraries expose a systemone call. For another small decision family, see our LFM3 guide.
How to update and try the new features
- Update Ollama from the app or your package manager, then confirm the version with
ollama --version. It should report 0.40.15 or later. - Run
ollamawith no arguments to open the new welcome launcher, or skip straight to a command. - Pull and test the coding model with
ollama run mimo-v2.6. - Connect it to an agent, for example
ollama launch claude --model mimo-v2.6. - On Apple silicon, try a routing decision:
ollama run autotrust-jev "My card was charged twice." --ask "Which team should handle this request?" --choice "billing=Charges and refunds" --choice "shipping=Deliveries and tracking". - If you use Kimi, run
ollama launch, choose Kimi Code CLI, and accept the offer to migrate your kimi-cli setup.
ollama launch changes
The launcher now covers more tools. Version 0.40.15 adds Kimi Code CLI, fills in Pi’s model list automatically so you can pick models inside Pi, and makes Claude Desktop’s model picker show Ollama’s actual model names. Version 0.40.14 made ollama launch claude apply each model’s output token limit, so Claude Code sessions on cloud models no longer stop at the 128K default cap. ollama launch codex now checks ChatGPT routing before configuring and reports a clear error when routing is unavailable. The desktop app can also retry after a ChatGPT disconnect instead of staying broken.
Limitations and upgrade notes
- In this release, autotrust-jev is supported only on Apple silicon, per the 0.40.15 notes.
- Clients that omit temperature or top_p on OpenAI-compatible endpoints now get the model’s packaged values rather than 1.0, so outputs may change after you upgrade. Set the values explicitly if you need the old behavior.
/v1/responsesnow keeps readable input content, including earlier assistant refusals, rather than flattening it.- Concurrent serving for recurrent and vision GGUF models is new, so measure memory use and throughput on your hardware before relying on it. For tuning GGUF engines directly, see our llama.cpp 0.9 guide.
Frequently asked questions
What is the latest version of Ollama?
As of November 8, 2026, the latest stable release on GitHub is Ollama 0.40.15, published November 3, 2026. It followed 0.40.14, which was published the previous evening Eastern time.
How do I run MiMo V2.6 in Ollama?
Update to Ollama 0.40.15 or later and run ollama run mimo-v2.6. The 9B model downloads at 6.6 GB to 8.7 GB, accepts text and images, and has a 256K-token context window.
What are AutoTrust decision models?
They are models that read text and return choices, yes or no probabilities, and ratings for questions you supply. The family comes in 9B, 35B and 27B-VL sizes, and the larger two can read images. In 0.40.15, autotrust-jev is supported on Apple silicon.
Can I use Kimi Code CLI with Ollama?
Yes. Starting with 0.40.15, ollama launch supports Kimi Code CLI, the successor to kimi-cli, and offers to migrate an existing kimi-cli setup.
Why do my OpenAI-compatible requests give different outputs after updating?
Since 0.40.14, Ollama’s OpenAI-compatible endpoints no longer force temperature and top_p to 1.0 when a client omits them. The model’s packaged sampler defaults apply instead, matching /api/chat.
Sources
- Ollama on GitHub: v0.40.15 release notes (November 3, 2026)
- Ollama on GitHub: v0.40.14 release notes (November 3, 2026 UTC)
- Ollama library: mimo-v2.6
- Ollama library: autotrust-jev
Image: Ollama logo by ParthSareen on behalf of Ollama, MIT License, via Wikimedia Commons. Ollama is a trademark of its owner. FreeAnswers is not affiliated with or endorsed by Ollama, Xiaomi or AutoTrust.
