MIMO-V2.6-PRO

MiMo-V2.6-Pro is a 1.02T-parameter Mixture-of-Experts (MoE) model with 42B active parameters. Text, images, video, and audio go in. Text comes out. The weights are MIT licensed, so you can download them and run them yourself.

Xiaomi published the model on September 22, 2026, and shipped more than a model. Alongside the weights sit a 44-page technical report, roughly 7,000 training task environments with their verifiers, an end-to-end reinforcement-learning (RL) training framework, small composable agent harnesses, and MiMo-V2.6-Distill-Qwen-9B, a 9.4B model distilled from the big one. The production RL run was streamed live while it happened.

That shape is worth knowing before you read the numbers. You can check every design choice against the shipped config file, rerun the environments that shaped the model's behavior, or start your own training from the distilled checkpoint. These chapters take one decision at a time and show what it costs.

Why the parameter count comes in two numbers

A dense model runs every parameter for every token. This one does not. Its feed-forward layers are split into 384 experts, and a router sends each token to 8 of them. The other 376 experts still have to be stored and served, because the next token may pick a different 8. That is the gap between 1.02T total and 42B active, and it is the gap between a model card and a serving bill.

The arithmetic is checkable against the shipped config, and it lands on the published numbers to rounding. Each expert is three matrices of 6144 by 2048 numbers, so 37.7M parameters. One layer holds 384 of them, and 69 of the 70 layers use experts.

the model (config.json + report Table 1):
  backbone     70 layers: 60 sliding-window, 10 global-attention
  experts      384 routed, 8 active per token, no shared experts
  params       1.02T total, 42B active per token
  encoders     MiMo-ViT 681M, audio 308M + 127M
  context      1M tokens, 131,072 max output, 152,576 vocab
  trained on   30T tokens (27T text + 3T omni), then agent mid-training

the release (September 22, 2026):
  weights      MiMo-V2.6-Pro-RL on HuggingFace, MIT license
  siblings     Flash 310B-A15B, Pro-UltraSpeed, Distill-Qwen-9B
  training     report, ~7k RL tasks with verifiers, RL framework

the meter (API pricing, September 2026):
  input        $0.435 per M tokens, $0.0036 on a cache hit
  output       $0.87 per M tokens

The shape of the model

Three inputs, one token sequence. The MiMo-ViT encoder turns images and video frames into visual tokens. Two audio encoders turn sound first into discrete tokens at 25 frames per second, then into one embedding per four frames. Text tokenizes as usual. Everything lands in a single sequence and passes through the same 70 layers.

We start with the window, because it sets the cache bill for everything else: 60 of the 70 layers read only the newest 128 tokens. From there the chapters cover the router and why it is frozen during training, the encoders and their token rates, the single mixed RL run, the grading loop that shaped the model's behavior, and finally the scores and the price.