MITmultimodalRuns locally309B (card)
MiMo-V2.6-Flash: Specs, License & How to Run
Xiaomi MIT Flash-RL MoE (card 309B/15B, 1M) with text, image, video, and audio.
- License
- MIT
- Parameters
- 309B (card) (active 15B (card))
- HF likes
- 582 (2026-10-25)
- First seen
- 2026-09-21
TL;DR. MiMo-V2.6-Flash is a multimodal model from Xiaomi MiMo. Parameters: 309B (card) (active 15B (card)).License: MIT. It is listed under ChatGPT, Closed LLM APIs.
| Developer | Xiaomi MiMo |
|---|---|
| Type | multimodal |
| Architecture | Sparse MoE (HF model_type mimo_v2; architecture MiMoV2ForCausalLM). Card: 48 layers (39 SWA / 9 GA), hidden 4096, 256 routed experts / 8 activated, no shared experts, 5-layer MTP drafter. Vision: 681M MiMo ViT. Audio: 308M AudioTokenizer + 127M patch encoder. safetensors.parameters sum to 310,756,322,688. |
| Parameters | 309B (card) · active 15B (card) |
| Modalities | text, image, video, audio |
| Languages | en, zh |
| Context | 1,048,576 tokens |
| License | MIT — Hugging Face cardData.license and the README YAML both say mit (opened 2026-10-25). There is no LICENSE file in the weight repo (raw /resolve/main/LICENSE returned 404 “Entry not found”). The companion code tree XiaomiMiMo/MiMo is Apache-2.0 and is not a substitute for the weight grant. Not legal advice. |
| Hardware | Not yet verified — Card usedStorage is 177,759,177,245 bytes (~166 GiB). Peak VRAM is not independently measured. README points at an SGLang cookbook with --tp 8. (not independently tested) |
| Hugging Face likes | 582 (as of 2026-10-25) |
| Model card | https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL |
| First seen | 2026-09-21 |
MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of Xiaomi’s MiMo-V2.6 series. Official hosts we opened on 2026-10-25: the ungated card XiaomiMiMo/MiMo-V2.6-Flash-RL (created 2026-09-21T15:39:51Z; last modified 2026-10-09T06:30:29Z; 582 likes / 90,978 downloads), the README (YAML license: mit), the architecture figure, the blog MiMo-V2.6, and the companion tree XiaomiMiMo/MiMo (2,394 stars, Apache-2.0). A LICENSE path on the weight repo returned 404.
This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT.
The README positions Flash-RL as a native omnimodal model (text, image, video, audio) with a 1M-token context. Hugging Face pipeline_tag is text-generation. Those sentences stay vendor.
What we verified (2026-10-25)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL (not gated) | HF API |
| Created / last modified | 2026-09-21T15:39:51Z / 2026-10-09T06:30:29Z | same |
| Likes / downloads | 582 / 90,978 | same |
| License tag | mit (card + README YAML) | HF cardData |
| LICENSE file in repo | Missing (HTTP 404) | HF raw |
| Pipeline tag | text-generation | HF API |
| Architecture | mimo_v2 (MiMoV2ForCausalLM) | card + HF config |
| Params (card) | 309B total / 15B activated | card |
| safetensors.parameters | 310,756,322,688 | HF API |
| usedStorage | 177,759,177,245 bytes (~166 GiB) | same |
| Context | 1,048,576 tokens | card + max_position_embeddings |
| Experts (config) | 256 routed, 8 active, no shared | HF config |
| Sibling named on the card | MiMo-V2.6-Pro-RL (separate host) | README |
We keep the README’s 309B / 15B quote and do not invent a reconciliation with the 310.8B safetensors sum.
How to run (from the card, not tested)
README: follow the SGLang MiMo cookbook; example flags include --trust-remote-code, --model-path XiaomiMiMo/MiMo-V2.6-Flash-RL, --tp 8. Docker image named: lmsysorg/sglang:latest. Hosted Xiaomi MiMo Studio / API / Desktop are product SKUs. forkalt has not served the shards.
License
Copied tags: Hugging Face cardData.license mit; README front matter license: mit. No LICENSE file was present on the weight repo on 2026-10-25. The GitHub tree XiaomiMiMo/MiMo is Apache-2.0 (code, not the Flash-RL shards). Not legal advice.
Card table (self-reported)
The README prints a vendor table. Flash-column rows we copied:
| Benchmark | MiMo-V2.6 Flash (card) |
|---|---|
| DeepSWE v1.1 | 67.9 |
| AutomationBench v1.0.6 | 52.3 |
| Terminal Bench 2.1 | 87.6 |
| OSWorld-Verified | 80.8 |
| CyberGym | 95.1 |
Other columns on that table name Pro, V2.5 Pro, Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5. Those stay vendor. forkalt has not reproduced any row.
What we did not invent
- VRAM and tokens/s stay unmeasured.
- We do not treat the missing LICENSE blob as Apache-2.0 just because the code repo is.
- We do not rank MiMo against ChatGPT or GLM-5.3-Flash.
Related: GLM-5.3-Flash — GLM-5.3-Flash vs DeepSeek-V4.1-Flash, DeepSeek-V4.1-Flash, MiniMax M3.1 Flash, Kimi K3, vLLM. Self-hosted UIs: Open WebUI, LibreChat, Jan. Pixel-editor traffic still lands on PhotoCraft vs GIMP — a different job.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| DeepSWE v1.1 (card, Flash column) | 67.9 (self-reported) | https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL |
| AutomationBench v1.0.6 (card, Flash column) | 52.3 (self-reported) | https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL |
| Terminal Bench 2.1 (card, Flash column) | 87.6 (self-reported) | https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL |
Deploy options
FAQ
Are the MiMo-V2.6-Flash weights public?
Yes on the host we opened. XiaomiMiMo/MiMo-V2.6-Flash-RL is ungated (created 2026-09-21T15:39:51Z; last modified 2026-10-09T06:30:29Z; 582 likes / 90,978 downloads in the 2026-10-25 API check). Flash-RL is the published Flash checkpoint; a separate Pro-RL card exists and is not this page.
Is the license a real MIT file?
The card and README YAML say mit. A LICENSE blob is not in the weight repo (404 on 2026-10-25). We copy that card tag and do not invent a file. The XiaomiMiMo/MiMo GitHub tree is Apache-2.0. Not legal advice.
Does it replace ChatGPT?
No sourced bake-off. The README table names Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5. Those rows stay vendor / self-reported.
How big is it?
Card: 309B total / 15B activated, 1M context. safetensors.parameters sum to 310,756,322,688. usedStorage 177,759,177,245 bytes. We keep the README 309B/15B quote and do not invent a reconciliation.
Sources & verification
- https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL
- https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/README.md
- https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL/blob/main/assets/architecture.png
- https://mimo.xiaomi.com/mimo-v2-6
- https://github.com/XiaomiMiMo/MiMo
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.