MiMo · model № 006
MiMo-V2.6-Pro
A 1.02T-parameter model that reads text, images, video, and audio, then trains every skill in one mixed reinforcement-learning run
total params
1.02T
384 experts on the shelf
active / token
42B
8 experts plus attention
context
1M
60 layers keep 128 tokens
price in / out
$0.435/$0.87
cache hits at $0.0036
Chapters
- 1Meet MiMo-V2.6-ProWhat shipped on release day, what the model does, and why the parameter count comes in two numbers.
- 2The 128-Token WindowSixty of seventy layers read only the newest 128 tokens. The other ten read everything, and the cache bill follows the split.
- 38 of 384Each token visits eight experts out of 384. The report shows what happened to the other 376 when the router was left trainable during RL.
- 4Ears and EyesHow a photo, a video clip, and a voice note become tokens in one shared sequence, and what each input costs.
- 5You Only RL OnceOne mixed reinforcement-learning run trained every capability at once: 30 steps, 25K trajectories per step, one $2.6M bill.
- 6Grading, and CheatingPass and fail cannot rank passing solutions, so grading got its own compute budget and an agent whose job is to break the tests first.
- 7Does It Deliver?Where the scores land against open and closed rivals, what the speed measurements say, and which claims still rest on in-house benchmarks.