Introduction

Xiaomi just opened the box on MiMo-V2.6 — and it is not a quiet weights dump. Today’s official note (September 22, 2026) ships two native omnimodal models, Pro and Flash, with MIT-licensed checkpoints on Hugging Face, a Distill-Qwen-9B research starter, and the RL harnesses Xiaomi used to push them. If you care about agents that can see, hear, and keep a million tokens of working memory, this is the open drop of the day.

What shipped

MiMo-V2.6 is a family, not a single file:

  • MiMo-V2.6-Pro-RL — flagship sparse MoE (model card: ~1.02T total / ~42B activated), 1M context, text + image + video + audio understanding
  • MiMo-V2.6-Flash-RL — efficiency twin (~309B total / ~15B activated on the card), same 1M context and modality set
  • MiMo-V2.6-Distill-Qwen-9B — a 9B SFT agent checkpoint on Qwen3.5-9B, meant as a community springboard for agentic RL

Collection home: XiaomiMiMo/mimo-v26. Cards point to SGLang / vLLM cookbooks; this is multi-GPU territory, not a weekend laptop toy.

Important boundary for creators: the public write-ups emphasize omnimodal understanding and agent action (computer use, Blender control, tool loops). They do not claim native Flux-style image/video generation from the open weights. Treat MiMo as a long-horizon multimodal brain that can steer creative tools — not as a drop-in replacement for MiniMax H3 or Qwen-Image.

Why the RL story matters

Xiaomi frames V2.6 as a bet on scaled reinforcement learning toward self-improvement: grow RL compute, environment diversity, and grader compute together so the model keeps stretching through exploration and feedback.

Highlights from the official release and model cards:

  • Live RL run under ~6 days, ~30 steps each for Pro and Flash, on the order of ~750,000 trajectories
  • Large asynchronous updates (cards cite 1,568 prompts × 16 rollouts per step, billions of tokens per update)
  • One mixed RL batch across coding, general agents, visual work, and cybersecurity — “you only RL once”
  • Groupwise agentic grading so rewards can rank passing solutions, not just pass/fail
  • Open follow-through: technical report, training environments, RL code, and lightweight harness pieces so others can reproduce

Vendor ranking claims (AA Intelligence Index, comparisons to Kimi / Qwen / Claude / GPT) are useful as Xiaomi’s scoreboard — treat independent benches as the long game.

Creator-flavored demos

Beyond SWE and terminal benches, Xiaomi shows MiMo stretching into creative and spatial loops:

  • Vibe World — natural language into interactive 3D / game-scene construction with multi-agent split of scene, logic, and visual checks
  • Blender modeling from text or reference images into assets aimed at animation, print, or games
  • Computer Use Agent — GUI understanding plus office/productivity actions with visual feedback
  • Design Arena–style front-end / PPT / SVG / video / music orchestration (Pro positioned near strong closed models on Xiaomi’s design board)

Again: the model is the planner and tool-user. Your favorite gen stack still paints the pixels.

How to try it

  1. Weights: pull Pro-RL, Flash-RL, or Distill-Qwen-9B from the HF collection above (MIT on the listed RL cards).
  2. Serve: follow Xiaomi’s SGLang MiMo cookbook (or the vLLM MiMo path) — expect serious GPU budgets for Pro/Flash.
  3. API: Xiaomi MiMo Open Platform keeps V2.5 pricing; API ids are lowercase mimo-v2.6-pro, mimo-v2.6-flash, and mimo-v2.6-pro-ultraspeed.
  4. Desktop: MiMo Desktop + membership ship with the models; UltraSpeed mode is pitched for low-latency interaction.

Original Source

https://mimo.mi.com/docs/en-US/news/latest/v2-6

https://huggingface.co/collections/XiaomiMiMo/mimo-v26

Conclusion

Open omnimodal agents with a million-token context used to be a closed-lab flex. MiMo-V2.6 puts Pro, Flash, a 9B distill, and the RL scaffolding on the table under MIT — with clear demos from Blender to computer use. Just keep the modality story honest: this drop is about understanding, memory, and agency, not replacing your image/video generators. Bookmark the collection, spin Distill-Qwen-9B if you want a research on-ramp, and watch how the community stress-tests those million-token tool loops next.

—Aurelia ♡