divinetribe/MiMo-V2.6-9B-MLX-4bit

MiMo-V2.6-Distill-Qwen-9B — Vision + Code (4-bit, MLX)

4-bit MLX quantization of Xiaomi's MiMo-V2.6-Distill-Qwen-9B, with the vision tower kept intact so it both reads images and writes code on Apple Silicon. 5.059 bits per weight (mixed; sensitive layers kept higher), 5.6 GB, group size 64.

4-bit is the smallest build, for 8 GB machines; the bf16 version is the full-quality one and the 8-bit version is the middle. This 4-bit build was put through the same pose-reading vision check as the others before publishing and still read the figure correctly, so the shrink held its sight.

MiMo-V2.6-Distill-Qwen-9B is a 9B agentic model from the Xiaomi MiMo team, built by supervised fine-tuning of Qwen3.5-9B on MiMo-generated data covering coding, general agent tasks, visual coding, and cybersecurity.

What it is

• Base: Qwen/Qwen3.5-9B (Qwen team).
• Source checkpoint: XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B (MIT).
• Architecture: Qwen3_5ForConditionalGeneration — a qwen3_5_vision tower feeding the qwen3_5_text decoder. 9.4B parameters.
• This repo: 4-bit MLX weights (group size 64, mixed precision), vision kept, for mlx-vlm.
• Other sizes: bf16 (full quality) · 8-bit.

Running it (MLX)

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("divinetribe/MiMo-V2.6-9B-MLX-4bit")
prompt = apply_chat_template(processor, model.config,
                             "What is in this image?", num_images=1)
print(generate(model, processor, prompt, image=["photo.jpg"], max_tokens=256))

Text-only works with num_images=0 and no image= argument.

Worked vision example — pose-match QC

A real frame from a motion game (a translucent target avatar drawn over a live camera scene) was fed to this 4-bit model, asking whether the drawn figure is a correct human pose and whether any limbs are extra or wrong. The model's unedited on-device answer:

The drawn figure is a single translucent human silhouette with two legs and two arms, so there are no extra or missing limbs. The arms hang down at the sides with elbows slightly bent, and the knees are bent with the legs in a narrow, slightly spread stance. The pose appears anatomically plausible, though the head is a simple rounded shape and the figure is overlaid on the real boxer, making the exact joint alignment somewhat ambiguous.

Even at 4-bit it is grounded in the actual pixels: it reads the translucency of the overlay, the single figure, the correct limb count (no extra or missing limbs — the exact QC signal), and the bent knees, and it correctly notices the avatar is overlaid on the real boxer. Like the bf16 and 8-bit builds it placed the arms low rather than up in a guard, so that is a trait of the model, not of the quantization. The same vision path writes alt-text for product and UI images and reads screenshots back into code.

Benchmarks

The single most reputable public benchmark that genuinely applies here is SWE-bench Verified — a human-validated set of real GitHub issues with hidden tests, the standard yardstick for coding agents. In the MiMo-V2.6 technical report, this SFT checkpoint scores 61.1 (avg@3) on SWE-bench Verified, ahead of its Qwen3.5-9B base at 60.0. Other numbers in that report (MiMo Code, MiMo Visual Coding, etc.) are the authors' internal evaluation sets, so treat those as the authors' own rather than an external leaderboard. 4-bit trades some accuracy for size, so expect the exact scores to move a little; the numbers here are quoted from the source card and report, not invented.

Credits

Enormous thanks to the people whose work this conversion simply repackages:

• The Xiaomi MiMo team — they built and released MiMo-V2.6-Distill-Qwen-9B under MIT: model · org. Citation: MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement, Xiaomi MiMo Team, 2026.
• The Qwen team at Alibaba — for the Qwen3.5-9B base and the Qwen3.5 vision-language architecture.
• The SWE-bench authors — Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, Karthik Narasimhan, and the OpenAI team behind the Verified split — for the benchmark quoted above: swebench.com.
• Prince Canuma and the mlx-vlm project, and the Apple MLX team, for the tooling that makes vision-language models run natively on Apple Silicon.

The quantization to 4-bit MLX is the only work done here. All model quality is theirs.

---

Part of Claude Code Local

This model is one of the fighters in Claude Code Local (3.2k★), which runs Claude Code 100% on-device on Apple Silicon through an MLX-native Anthropic-API server. Not sure which local model to run as an agent? Check the Agent-12 local agent leaderboard: real agent tasks, judged by the filesystem, same hardware for every row.

Built by Matt Macosko in Humboldt, CA. Open to work on local-AI and Apple Silicon inference: matt@ineedhemp.com.