MiMo V2.6 9B on a Mac: sizes, Ollama tags and speed

Sizes and memory ↓Published speed ↓Model-fit calculator →

Which MiMo V2.6 model runs on a Mac?

Xiaomi put MiMo V2.6 on Hugging Face on 21 September 2026 in three sizes:

The 9B is not a new architecture. Xiaomi fine-tuned Qwen3.5-9B on outputs from the larger MiMo models. It keeps Qwen’s 262K context and takes images as well as text. On Xiaomi’s own card it scores 44.6 on SWE Pro against 32.0 for the base Qwen3.5-9B, a large jump for a distill, and not one I have checked.

MiMo V2.6 9B file sizes and memory

The Ollama q4_K_M tag is about 0.8 GB larger than the GGUF of the same name, close to the size of the vision projector, so it probably includes it. The “Mac” column is my half-of-memory planning rule, which leaves room for a long context. The calculator shows how the KV cache grows if you actually use the 262K window.

How fast it runs on Apple silicon

The clearest single run I found is from Onticpost, updated 25 October (community). On an M4 Pro with 48 GB, using the ggml-org Q8_0 file in LM Studio (llama.cpp Metal runtime 2.47.0) at a 16K context, they measured:

• Median decode of 24.09 tokens a second over five runs.
• Median prompt processing of 328 tokens a second.
• 10.80 GB resident memory.

They note the machine was already swapping during the runs, so a clean machine may do better. Aggregated oMLX runs on Token Atlas (secondary, collected from community submissions) show a 4-bit MLX build at 84 to 88 tokens a second on an M4 Max and about 40 to 45 on an M4 Pro at short contexts. Those use a different engine and a smaller quant, so do not compare them directly with the Q8_0 number.

For scale: a 9B at 24 tokens a second is fast enough to read along with. If you want a bigger model in the same memory, my 24 and 48 GB list compares the other recent releases.

Running MiMo V2.6 in Ollama, LM Studio and llama.cpp

• Ollama. ollama run mimo-v2.6 pulls the library model. MiMo support was merged in PR 19400 on 3 November and first shipped in the 0.40.15-rc0 pre-release. The current stable build is 0.40.14. I have not confirmed whether 0.40.14 loads the tag, so if the pull fails, try the pre-release.
• LM Studio. Search for ggml-org/MiMo-V2.6-Distill-Qwen-9B-GGUF or bartowski’s repo. Skip the mmproj file unless you need images.
• llama.cpp. Because the 9B is a Qwen 3.5 model underneath, any recent build loads it. Pass --mmproj for image input.

ollama run mimo-v2.6

LM Studio vs Ollama covers which app fits which workflow.

The license is not plain MIT

The Hugging Face license field says MIT, and I found no LICENSE file in the repo. The model card then lists use restrictions, including no military use. Plain MIT has no such clauses. For personal use this changes nothing. For a product, read the card in full and get advice before you rely on the MIT tag.

FAQ: MiMo V2.6 on a Mac

Related local notes

• Which new local models fit a 24 GB or 48 GB Mac
• Running Jais 3 locally
• MiniCPM-V 4.7 on a Mac
• Qwen 3.8 27B on my M4 Max

Sources and what I did not test

I have not run MiMo V2.6. The speed figures are community runs, the scores are Xiaomi’s, and the sizes were read from the download pages on 4 November 2026.

Sources

• Xiaomi: MiMo-V2.6-Distill-Qwen-9B model card ↗
• Xiaomi: MiMo-V2.6-Flash-RL ↗
• Xiaomi: MiMo-V2.6-Pro-RL ↗
• bartowski: MiMo 9B GGUF files ↗
• Ollama library: mimo-v2.6 ↗
• Ollama PR 19400: MiMo support ↗
• Onticpost: MiMo 9B on a 48 GB M4 Pro ↗
• LLMCheck: local vision models for Mac by RAM ↗
• Token Atlas: collected oMLX runs ↗

Suhail Mohebi · published 4 November 2026 · the cover is a generated illustration