Xiaomi Releases MiMo-V2.6 Weights Under MIT, a 1.02-Trillion-Parameter Pro Model

Xiaomi released MiMo-V2.6 in the week of Sept. 21, with an MIT licence that permits commercial use. The flagship, MiMo-V2.6-Pro, is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per token, according to its Hugging Face model card.

Sources disagree on the exact day. Times of AI gives Sept. 21 and the AI Frontier Post gives Sept. 22.

The two checkpoints

The release has a Pro and a Flash model. Both take text, image, video and audio and have a 1 million token context window, according to the model cards. Times of AI also lists a third variant, Pro-UltraSpeed, which it says trades quality for speed. That variant was not among the cards checked for this piece.

Spec	Pro	Flash
Total parameters	1.02T	309B
Active per token	42B	15B
Licence	MIT	MIT
Context	1M tokens	1M tokens

Pro uses 384 routed experts with 8 active per token and a 70-layer backbone, per Times of AI. Flash has a 48-layer backbone with sliding window attention, per its card. The weights are on Hugging Face and ModelScope, and the AI Frontier Post says they are ungated. The cards list SGLang and vLLM for serving. Neither card gives a memory requirement, so the hardware needed for a trillion-parameter checkpoint is left open.

Vendor benchmarks and the independent score

The benchmark figures on the cards are Xiaomi's own. The Pro card lists 71.9 on DeepSWE v1.1, 89.9 on Terminal Bench 2.1 and 94.0 on CyberGym. The Flash card lists 67.9 on DeepSWE, 73.6 on Toolathlon and 95.1 on CyberGym. Flash scoring higher than Pro on CyberGym is in Xiaomi's own tables, and the cards give no reason for it.

The one third-party figure is from Artificial Analysis. Times of AI reported a Pro score of 46 on the Intelligence Index v4.3.2, level with Grok 4.7 at its xhigh setting. It put GLM-5.3 at 45 and Kimi K3 at 44. The AI Frontier Post gave 46.32 on v4.3, so the two outlets label the index version differently.

1. MiMo-V2.6-Pro46 points
2. GLM-5.345 points
3. Kimi K344 points

Source: Times of AI, citing Artificial Analysis, accessed 2026-09-28

The AI Frontier Post added that, in Xiaomi's testing, Pro trails Claude Opus 5 on some software-engineering tasks, 71.9 to 74.0. It said independent evaluation of the release is still pending.

The training cost disclosure

Xiaomi livestreamed the reinforcement-learning runs for about six days. Both outlets report a cost of roughly $850,000 for Flash and $2.62 million for Pro, across about 750,000 trajectories and more than 7,000 task environments. Xiaomi has promised to open the environments and training framework under MIT.

API pricing for Pro is $0.435 per million input tokens and $0.87 per million output tokens, per Times of AI and the AI Frontier Post.

Licence details matter on releases like this. The site's report on the Hemmingway-1 licence mismatch covers a card and a licence file that disagreed. For MiMo-V2.6 the cards and both outlets say MIT. Nvidia's Nemotron diarization model was also trending on Hugging Face this week.

The sources checked give no date for the release of the environments.