Bottom line up front
MiMo-V2.6-Flash is a sparse mixture-of-experts model with about 309 billion total parameters and roughly 15 billion active per token, released by Xiaomi under the MIT licence alongside its larger Pro sibling. Xiaomi documents natively omnimodal input covering text, image, video, and audio, a 1M-token context, and unchanged V2.5 pricing on its own endpoint. Because the weights are MIT-licensed and the technical report and reinforcement-learning environments are published, self-hosting is a genuine option — the constraint is hardware rather than permission.
A 309B model that computes like a mid-sized one
Xiaomi describes the MiMo-V2.6 series as natively omnimodal, with Pro as the most capable model and Flash as the balance of intelligence, efficiency, and cost. Third-party indexes record Flash as roughly 309 billion total parameters with about 15 billion active per token, against Pro at roughly 1.02 trillion total and 42 billion active. Active parameters drive per-token compute, so Flash is positioned as the cost-efficient member of the family.
For a team, the practical reading is that the capability ceiling comes from the full parameter pool while the compute bill looks like a mid-sized model. That is the same trade DeepSeek and NVIDIA make in their sparse models, and it means comparing models on active parameters tells you more about cost than comparing total parameters.
- Record total and active parameters separately for every candidate.
- Use active parameters as the cost and latency signal.
- Confirm modality requirements before treating omnimodality as a requirement.
Omnimodal is a claim to test, not a checkbox
Xiaomi states the V2.6 series takes text, image, video, and audio input. That is a broad claim, and the useful question is which of those inputs your workflow actually needs and whether the hosted endpoint you call accepts them. A model being omnimodal does not mean every provider serves every modality, or that the quality across modalities is equivalent.
Test the specific modalities you need on the specific endpoint you intend to use. A model card stating a capability is evidence that the capability exists in the weights; it is not evidence that your access path supports it at a quality you accept.
- List the modalities your workflow actually requires.
- Verify each one against the hosted endpoint, not just the model card.
- Test quality per modality rather than assuming it transfers.
MIT weights make self-hosting a real option
Xiaomi released the MiMo-V2.6 weights under the MIT licence, alongside a technical report, the reinforcement-learning task environments, the RL framework, and the agent harnesses. Third-party reporting records more than 7,000 RL task environments published, which is a substantial transparency claim for a model of this class.
MIT permits commercial use, modification, and redistribution with attribution, so the permission is not the barrier. Hardware is: a 309-billion-parameter model needs substantial memory even with only a fraction active per token. That is a capacity question for your own infrastructure, and one the model card cannot answer for you.
- Locate and read the licence for the exact checkpoint you intend to run.
- Record the model version and the date you checked the licence.
- Estimate memory from the checkpoint size plus context and KV-cache overhead.
Read the published price and the free tier together
Xiaomi's own endpoint lists MiMo-V2.6-Flash at roughly two and a half cents per million cached input tokens, fourteen cents per million uncached input, and twenty-eight cents per million output, with pricing unchanged from V2.5. Those are the developer's own figures for its own platform; third-party hosts set their own, and gateway prices vary widely for the same model.
MiMo-V2.6-Flash is also reachable at zero cost on some gateways. That is a subsidy over a real cost, with rate limits and no continuity commitment. Where the MIT licence applies, the second route — running it yourself — converts a subsidy you do not control into a cost you can forecast, subject to the hardware question above.
- Compare the same model across providers rather than assuming one price.
- Record cached and uncached input prices separately; they differ by an order of magnitude.
- Budget for a paid tier or a self-hosted path for anything that must keep running.