AI tools ·

Xiaomi opens MiMo-V2.6-Pro weights under MIT — the new top open-weights model

Xiaomi published full, ungated, MIT-licensed weights for MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL on Hugging Face on September 21, with no launch fanfare — the Flash checkpoint appeared around 15:39 UTC, the Pro weights seconds later. Pro is a 1.02-trillion-parameter sparse mixture-of-experts (42B active per token) with a 1-million-token context window and natively omnimodal input (text, image, video, and audio in; text out). The Flash sibling is a ~309B-total / 15B-active variant. Artificial Analysis ranks MiMo-V2.6-Pro as the world's highest-scoring open-weights model with an Intelligence Index of 46, ahead of DeepSeek V4.1 Flash.

The release ships with weights, a technical report, and deployment instructions. Xiaomi has committed to also publishing the training environments and RL code, but those were not in the repos at the time of reporting — one thing the coverage conflicts on, so don't download expecting the full training stack yet. There is also a Pro-UltraSpeed variant claiming up to 20× faster output at the same quality. Xiaomi-reported benchmarks include DeepSWE v1.1 scores of 71.9 (Pro) and 67.9 (Flash) and CyberGym scores of 94.0/95.1 — vendor-reported, with no independent rerun yet.

Press-reported API pricing holds at V2.5 levels: Pro at $0.435 per million uncached input tokens and $0.87 output; Flash at $0.14 and $0.28. One caveat: Xiaomi's own blog still shows no published price list for the new generation, so treat those figures as Xiaomi-stated rather than confirmed on a vendor price sheet.

Why it matters

This is the open-weights counterweight to the closed frontier releases arriving the same week: MIT-licensed weights you can download free and self-host, at the top of the open-weights leaderboard. For anyone who can't ship data to a US API — or just wants a frontier-grade coding agent on their own hardware — this is the new default download.

Key facts

  • Weights published September 21 on Hugging Face (XiaomiMiMo/MiMo-V2.6-Pro-RL and -Flash-RL), ungated, MIT license
  • Pro: 1.02T total / 42B active sparse MoE; Flash: ~309B total / 15B active; both 1M-token context, omnimodal input
  • Artificial Analysis: Intelligence Index 46 — top-ranked open-weights model, ahead of DeepSeek V4.1 Flash
  • Pro-UltraSpeed variant claims up to 20× faster output at the same quality
  • Xiaomi-reported benchmarks: DeepSWE v1.1 71.9 (Pro) / 67.9 (Flash); CyberGym 94.0 / 95.1 (vendor-reported, no independent rerun)
  • Press-reported API pricing: Pro $0.435/$0.87, Flash $0.14/$0.28 per million input/output tokens (not confirmed on a vendor price sheet)

Sources

Spot an error? We correct quickly and note it. Contact us via the contact page.

← More AI news