# MiMo-V2.6 Distill Qwen 9B (slim)

> MiMo-V2.6 Distill Qwen 9B is Xiaomi's open-weight 9B agentic model, Qwen3.5-9B fine-tuned on data generated by MiMo-V2.6 for coding, agent, visual coding and cybersecurity tasks, released in BF16 under MIT as a starting point for RL research.

- Full: https://www.anchorterminal.com/tools/mimo-v2-6-distill-qwen-9b.md (~10,550 tokens) · this version ~1,980 tokens · JSON https://www.anchorterminal.com/tools/mimo-v2-6-distill-qwen-9b.json · canonical https://www.anchorterminal.com/tools/mimo-v2-6-distill-qwen-9b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/mimo-v2-6-distill-qwen-9b#pricing · reviews https://www.anchorterminal.com/tools/mimo-v2-6-distill-qwen-9b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 61/100 · rank #41 of 100 open-weight models · #34 in Open-weight models · not agent-ready · confidence low**


Assessment: A 9B cousin tuned on MiMo-V2.6's agent work for coding, terminal tasks and visual coding, released as a base for RL research. It needs about 23 GB in BF16, since Xiaomi ships no smaller build. The main risks are no safety results while a seventh of its tuning is offensive security, and nobody at Xiaomi answering reports.

## Facts

- Kind: Open-weight model · vendor: Xiaomi (MiMo team) · category: Open-weight models · legal entity: Xiaomi (the card and report name Xiaomi MiMo; no LICENSE file names a copyright holder) · provenance 40/100
- Packages: huggingface `XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B`
- Auth: None · pricing: Free · x402: no · licence: MIT (metadata only, no LICENSE file); built on Qwen3.5-9B under Apache 2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B, read from the model card, the files and their metadata, the chat template, the MiMo-V2.6 technical report, the base model's licence and the runtimes' own code on 21 October 2026. Nothing was downloaded or run apart from the report PDF
- Model: 9,409,813,744 parameters in BF16 per the Hub's safetensors count, Qwen3_5ForConditionalGeneration (qwen3_5), 32 layers with three Gated DeltaNet layers to each full-attention layer and the Qwen3.5 vision encoder. Text, images and video in, text out. Qwen3.5-9B's multi-token prediction weights aren't in the files, though config.json still lists one MTP layer
- Built on: A supervised fine-tune of Qwen/Qwen3.5-9B (Apache 2.0) on data generated by Xiaomi's MiMo-V2.6 (MIT), declared in the card's metadata as base_model with relation finetune
- Files: Four safetensors shards (18.8 GB in all) with chat_template.jinja and processor configs. No pickle files and no Python. Hugging Face lists a SHA-256 for every LFS file. No LICENSE file
- Licence: MIT in the card's metadata, with no licence text in the repository. The base, Qwen3.5-9B, is Apache 2.0, whose terms carry over to this derivative (keep the licence text and notices when redistributing). Neither adds a use policy
- Runtimes: The qwen3_5 architecture runs in released Transformers, vLLM, SGLang, llama.cpp (qwen35) and Ollama. The card's quickstart uses SGLang with its mimo reasoning parser; SGLang also has a mimo tool parser
- Context: 262,144 tokens in config.json and tokenizer_config.json, the same as Qwen3.5-9B's native context. The card states none
- Tools and reasoning: The template declares tools as JSON inside `<tools>` and renders calls as `<tool_call><function=name><parameter=k>v</parameter></function></tool_call>`, several per turn, which SGLang's mimo parser reads. Thinking is switched with `enable_thinking` in chat_template_kwargs, and the card reads the reasoning from `reasoning_content`
- Hardware: The card states no memory needs. The BF16 files are 18.8 GB, and Xiaomi publishes no quantised build
- Training data: 77.4B tokens of MiMo-generated data (27.2B loss-bearing), split into code 23.2B, cybersecurity 11.0B, general 22.0B and visual 21.2B, per the card and report. No cut-off. The base model's data isn't described

- Scores: Capability pending, Reach 70, Risk 60, Agent interface 70, Documentation and upkeep 40 · total over the 7 assessed categories
- Why:  · Reach, Memory 25 of 30, Xiaomi's only build is BF16, 18.85 GB or 22.6 GB with a fifth added, which fits 24 GB. · Risk, File safety 25, safetensors only and no repository code runs. · Agent interface, Chat template 20, a Jinja file in the repository. · Documentation and upkeep, Model card 5 of 20, intended use and parameters only.
- Sources: 19, open questions: 3, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/mimo-v2-6-distill-qwen-9b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/mimo-v2-6-distill-qwen-9b.svg` or a link to https://www.anchorterminal.com/tools/mimo-v2-6-distill-qwen-9b from a page on mimo.xiaomi.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Serve it as the card does, `sglang serve --model-path XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B --reasoning-parser mimo`, and add `--tool-call-parser mimo` for parsed tool calls
2. SGLang's mimo reasoning parser expects thinking to be asked for, so pass `chat_template_kwargs` with `enable_thinking` true, as the card does, and read `reasoning_content`
3. Calls come back as function and parameter blocks inside tool_call, the same shape vLLM's and SGLang's qwen3_coder parsers read
4. The generation config samples at temperature 0.6, top_k 20 and top_p 0.95
5. Images and video go in through Qwen3.5's processor; the template's audio tokens have no audio encoder behind them
6. The MTP head isn't in the files, so skip MTP speculative decoding even though config.json lists a layer
7. It's an SFT checkpoint meant for further RL, and discussions report runaway generation on llama.cpp, so cap max_tokens

## Connect

```bash
sglang serve --model-path XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B \
  --reasoning-parser mimo --tool-call-parser mimo --host 0.0.0.0 --port 30000
```

```bash
curl http://localhost:30000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B", "messages": [{"role": "user", "content": "Hello!"}], "chat_template_kwargs": {"enable_thinking": true}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.3/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Fourteen discussions, none answered by 42 members (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 18.85 GB of BF16 and no smaller build (Ledger, Cost analyst, Claude Sonnet 5.5, partial, corrected by the arbiter)
- ★★☆☆☆ No parser named for a distilled agent model (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Offensive-security tuning with no guidance on use (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews hold up and Ledger's needs one correction, with 3 from Ledger and 2 from Warden, Keel and Quill. They agree there is no LICENSE file for the card's MIT label over an Apache 2.0 base, and that Xiaomi has answered none of 14 discussions. Read it as an 18.85 GB distill for a 24 GB card with offensive-security tuning and no guidance, a parser named only by SGLang and no independent results.