MiMo-V2.6 9B in Ollama: Xiaomi's Coding Model for 16 GB Computers

Ollama 0.40.15, released on 3 November 2026, added MiMo-V2.6-Distill-Qwen-9B, a 9-billion-parameter model from Xiaomi MiMo for coding and agent work, under the MIT license. Its 4-bit download is 6.6 GB, which leaves about 9 GB on a 16 GB computer for the conversation, the model runner and everything else you have open. Enter your own hardware in the calculator before downloading it. Xiaomi's published gains are in coding-agent and tool tasks; its figures do not cover ordinary chat.

What Xiaomi released

MiMo-V2.6-Distill-Qwen-9B is Qwen3.5 9B with further training on data generated by Xiaomi's larger MiMo models. Xiaomi describes it as covering coding, general agent tasks, coding from screenshots and security analysis, and says it is publishing this version as a starting point for open research.

The published context limit is what the model can accept, not what your computer can hold. Start with a short context and raise it only as far as your memory allows.

Download sizes

The default tag, mimo-v2.6, is listed at 6.6 GB to 8.7 GB: Ollama chooses between the two 4-bit versions. The 4-bit files are compressed (quantized) copies of the model. The quantization guide explains what that costs in answer quality.

How much memory it needs

The download is the starting point, not the total. Add conversation memory (KV cache) for the context length you set, memory for the model runner, and what the operating system and your other apps already use.

The RAM and VRAM calculator lists MiMo: open MiMo-V2.6 9B in the calculator. The link starts at 4,096 context tokens and leaves your hardware blank until you enter it. The calculator uses a 5.8 GB community GGUF file; Ollama's own 4-bit package is 6.6 GB, so add about 1 GB to the estimate if you install through Ollama.

As a guide to reading the result:

• 16 GB of system memory (RAM) or unified memory: the 6.6 GB file leaves about 9 GB for the conversation, the runner, the operating system and other apps. That is workable for short and medium contexts with other heavy apps closed. See the 16 GB setup guide.
• A graphics card with 8 GB of dedicated graphics memory (VRAM): the 6.6 GB file leaves about 1.4 GB on the card. Expect to keep the context short, or to let part of the model run from system RAM, which is slower.
• 8 GB of RAM with no graphics card: the file alone takes most of the memory. Choose a smaller model; the 8 GB setup guide lists options.

Run it in Ollama

You need Ollama 0.40.15 or newer. If Ollama is not installed, follow the Ollama installation guide.

ollama --version
ollama run mimo-v2.6

The second command downloads the model (6.6 GB to 8.7 GB) and opens a chat. To choose the smaller file yourself:

ollama run mimo-v2.6:9b-q4_K_M

Ollama's page also gives commands that connect the model to coding tools, for example:

ollama launch opencode --model mimo-v2.6

What Xiaomi's benchmarks show

These are Xiaomi's figures, comparing MiMo with Qwen3.5 9B. Local LLM Guide has not tested them.

Three things to keep in mind when reading the table:

• The gains are in agent work, not in every coding test. SWE Verified barely moves; the larger changes are in tasks that need many tool calls in a row.
• Some of Xiaomi's other results use its own test sets. Its model card marks four as internal evaluations. Nobody outside Xiaomi can check those.
• The scores come from Xiaomi's test setup, not from the 4-bit download. Xiaomi does not publish scores for the compressed files.

A fair test before you switch

If Qwen3.5 9B or a similar model already works for you, compare the two on your own work before changing:

1. Pick one small, reversible change in a scratch copy of a project, such as adding a test or renaming a function across two files.
2. Give both models the same instruction through the same coding tool.
3. Check whether each edit applies cleanly, whether the code still runs, and how many times you had to correct the model.

A model that chats well can still return edits a coding tool cannot apply. Using Continue, Cline or Aider with a local model shows how to set up that test, and choosing a local model for coding compares other sizes.

The license and Xiaomi's use notice

The model is released under the MIT license, which permits commercial use, modification and redistribution as long as the copyright and permission notice stay with copies.

Xiaomi's model card adds a section headed "Use Restrictions". It lists uses you agree to avoid, including unlawful or malicious purposes, military applications, malicious cyber activity, handling personal data without authorization, and letting the model take high-risk actions without human oversight. The same section then says the notice is an informational reminder that "does not amend the applicable model licence". Read both the license and the notice before using the model at work, particularly for security testing, which the model is trained for and the notice addresses.

Keep it local

ollama run mimo-v2.6 runs the model on your own computer after the download. Coding tools add their own connections: a tool may send code to its maker's service, fetch web pages or run commands. Check each tool's settings before opening a private project, and see the local AI privacy guide for what to look for.

Also in Ollama 0.40.15

• AutoTrust decision models, which return choices, yes-or-no probabilities and ratings instead of prose. The release notes say they run on Apple silicon in this release. Our guide to Ollama decision models explains how that kind of model is used.
• Vision models in GGUF format can now serve more than one request at a time.
• Running ollama with no arguments opens a welcome launcher.

Related reading

• MiMo-V2.6 9B: memory, downloads and a function-repair check
• Qwen3.5 9B: memory and setup
• Ripple-65B Preview: a much larger coding model added in Ollama 0.40.16
• What local LLM can my computer run?

Sources

• Ollama 0.40.15 release notes
• Ollama's MiMo-V2.6 page and its list of versions and sizes
• MiMo-V2.6-Distill-Qwen-9B model card: training, benchmarks, license and use notice