The pack
Xiaomi's MiMo-V2.6-Distill-Qwen-9B is Qwen3.5-9B fine-tuned by Xiaomi MiMo for coding, agent work, visual coding and security tasks. In MTPLX it is MiMo V2.6 Qwen 9B Optimized Speed.
The pack,
Youssofal/MiMo-V2.6-Qwen-9B-MTPLX-Optimized-Speed,
is 8.70 GB: the model at 6 bits with group size 64, the same recipe as
Qwen 3.5 9B Optimized Speed, with Xiaomi's vision tower in BF16. Xiaomi's
checkpoint has no draft head, so the pack uses the Qwen3.5-9B one. The served id is
mtplx-mimo-v26-qwen-9b-optimized-speed. The base model is Xiaomi's
MiMo-V2.6-Distill-Qwen-9B. Full credit
for the model goes to Xiaomi MiMo, and it is MIT licensed.
Xiaomi's model card reports it well ahead of the Qwen3.5-9B it started from:
| Benchmark (Xiaomi's model card) | Qwen3.5-9B | MiMo-V2.6-Distill-Qwen-9B |
|---|---|---|
| SWE Pro | 32.0 | 44.6 |
| Terminal Bench 2.1 | 27.0 | 37.1 |
| Toolathlon-Verified | 25.9 | 35.2 |
| SWE Verified | 60.0 | 61.1 |
The two models have the same architecture and size.
Speed and quality
MTPLX has not measured this pack's decode speed yet. Each forward pass does the same work as on Qwen 3.5 9B Optimized Speed, which has the same architecture and size, but the draft head accepts fewer tokens. What has been measured on MTPLX:
| Measurement | Result | Conditions |
|---|---|---|
| Decode speed | Not measured yet | |
| Draft acceptance, first position | 78.5 percent | Long game prompt, thinking on |
| Draft acceptance, second position | 56.5 percent | Long game prompt, thinking on |
| Top-1 agreement with Xiaomi's BF16 checkpoint | 97.3 percent | 19,265 tokens of code and prose |
| KL divergence to Xiaomi's BF16 checkpoint | 0.0054 | 19,265 tokens of code and prose |
| Peak memory while serving | 8.70 GiB | 15K-token context |
The draft head was trained on Qwen3.5-9B, so it accepts fewer drafts on MiMo. Depth 2 is the default, as on the Qwen 9B. The 6-bit pack is ten times closer to Xiaomi's BF16 checkpoint than a mixed 4-bit build of the same model.
Which Mac
In MTPLX 2.12.0 the app and the CLI read one catalog with the same recommendation rule. On M3, M4 and M5 Macs with 16 GB or more they list MiMo V2.6 Qwen 9B right before Qwen 3.5 9B.
| Mac | What the app and the CLI suggest |
|---|---|
| M3, M4 or M5 with 16 to 31 GB | Ternary Bonsai 2 27B first, MiMo V2.6 Qwen 9B second |
| M3, M4 or M5 with 32 to 255 GB | Qwen 3.8 27B Optimized Speed first |
| M3, M4 or M5 with 256 GB and above | Flash-Next Optimized Speed first |
| M3, M4 or M5 below 16 GB | Qwen 3.5 4B Optimized Speed first |
| M1 and M2 | The FP16 Qwen 3.5 9B stays the suggestion |
- 16 to 31 GB: the first suggestion, Ternary Bonsai 2 27B, is a 27B-class model in about the same memory.
- Agent clients: the pack card recommends 18 GB or more for agent clients such as OpenCode.
- Existing choice: a model you already picked stays selected.
Memory and download
The download is 8.70 GB. MTPLX plans the same context windows as for Qwen 3.5 9B Optimized Speed, and the measured peak was 8.70 GiB at a 15K-token context.
| Mac memory | Context window |
|---|---|
| 16 GB | 20,480 tokens |
| 18 GB | 45,056 tokens |
| 24 GB | 192,512 tokens |
| Largest window | 262,144 tokens |
- Recommended: 16 GB or more of unified memory, and 18 GB or more for agent clients such as OpenCode.
- SSD session cache: 16 GiB by default on Macs with 16 GB of RAM or less, and 24 GiB on Macs with up to 32 GB.
How to run it
Mac app: download MTPLX and pick "MiMo V2.6 Qwen 9B Optimized Speed". It needs MTPLX 2.12.0 or later.
Command line, with pip:
pip install mtplx
mtplx serve --model Youssofal/MiMo-V2.6-Qwen-9B-MTPLX-Optimized-SpeedOr with Homebrew:
brew install youssofal/mtplx/mtplx
mtplx serve --model Youssofal/MiMo-V2.6-Qwen-9B-MTPLX-Optimized-SpeedThe serving contract ships inside mtplx_runtime.json, and MTPLX reads it on load. Sampling is
temperature 0.6, top-p 0.95, top-k 20, Xiaomi's settings, and the draft depth is 2.
Siblings: Qwen 3.5 9B, the model MiMo started from, and Bonsai 2 27B, the MTPLX pack of Prism ML's Ternary-Bonsai-2-27B-mlx-2bit, a 27B-class model in about the same memory. The full list of changes is in the MTPLX 2.12.0 release notes.
How it is built
- 6-bit weights. The model is quantized to 6 bits with 64-weight groups, the recipe of Qwen 3.5 9B Optimized Speed.
- Vision tower. Xiaomi's vision tower stays in BF16.
- Draft head. Xiaomi's checkpoint has no draft head, so the Qwen3.5-9B multi-token-prediction head drafts ahead, and the full model verifies its drafts in one pass. Depth 2 is the default, as on the Qwen 9B.
Speculation in MTPLX is exact. Drafts are accepted with the probability-ratio rule plus residual resampling, so the output follows the model's own distribution at any temperature. The 6-bit quantization is measured against Xiaomi's BF16 checkpoint in the table above.
Questions
Can I run Xiaomi MiMo on a Mac?
Yes. Xiaomi's MiMo-V2.6-Distill-Qwen-9B runs on MTPLX 2.12.0 or later as MiMo V2.6 Qwen 9B Optimized Speed, an 8.70 GB pack for any Apple Silicon Mac with 16 GB of memory or more. Full credit for the model goes to Xiaomi MiMo, and it is MIT licensed.
Can a 16 GB Mac run MiMo V2.6 Distill Qwen 9B?
Yes. MTPLX plans a 20,480-token context window on 16 GB, 45,056 on 18 GB and 192,512 on 24 GB, and the measured peak was 8.70 GiB at a 15K-token context. The pack card recommends 18 GB or more for agent clients such as OpenCode.
How fast is MiMo V2.6 Qwen 9B on a Mac?
Its speed on MTPLX has not been measured yet. Each forward pass does the same work as on Qwen 3.5 9B Optimized Speed, which has the same architecture and size. The draft head was trained on Qwen3.5-9B, so it accepts fewer drafts on MiMo: 78.5 percent at the first position and 56.5 percent at the second, on a long game prompt with thinking on.
How do I run MiMo-V2.6-Distill-Qwen-9B on a Mac?
Download the MTPLX Mac app at mtplx.com/download and pick "MiMo V2.6 Qwen 9B Optimized
Speed". From the command line, run pip install mtplx or brew install youssofal/mtplx/mtplx,
then mtplx serve --model Youssofal/MiMo-V2.6-Qwen-9B-MTPLX-Optimized-Speed. Both routes need MTPLX
2.12.0 or later.
Is MiMo V2.6 Qwen 9B better than Qwen 3.5 9B for coding?
Xiaomi's model card reports it well ahead of the Qwen3.5-9B it started from: 44.6 against 32.0 on SWE Pro, 37.1 against 27.0 on Terminal Bench 2.1, 35.2 against 25.9 on Toolathlon-Verified and 61.1 against 60.0 on SWE Verified. The two models have the same architecture and size.
How close is the 6-bit MiMo pack to Xiaomi's original model?
Against Xiaomi's BF16 checkpoint over 19,265 tokens of code and prose, the 6-bit pack has a KL divergence of 0.0054 and picks the same top token 97.3 percent of the time. That is ten times closer than a mixed 4-bit build of the same model.