Xiaomi ships a 1T MIT-licensed model. Weights you own.

Xiaomi put a trillion-parameter omnimodal model on Hugging Face under an MIT license. MiMo-V2.6 Pro and Flash both shipped September 21 with open weights, 1M-token context, and the most permissive license in common use. For a small team, the license line is the news — not the parameter count.

What actually happened

The MiMo-V2.6-Pro-RL model card lists a sparse mixture-of-experts at 1.02T total parameters with 42B activated, 70 layers, 384 routed experts with 8 active per token, and a 1M-token maximum context. Modalities are text, image, video and audio.

The smaller MiMo-V2.6-Flash-RL is the one most teams would actually run: 309B total, 15B activated, 48 layers, 256 routed experts with 8 active, same 1M context, same four modalities. Fifteen billion active parameters is a serving cost you can reason about. A 9B distilled variant is published alongside them.

Both are MIT. Not a community license with a revenue threshold, not a research-only license, not a bespoke acceptable-use addendum. MIT — the same four-paragraph license on half the packages in your node_modules.

We are deliberately not quoting training costs or reinforcement-learning task counts circulating for this release. Those numbers appear in aggregator writeups we could not trace to Xiaomi's own materials, so we are leaving them out rather than repeating them.

Why an MIT-licensed model matters for your business

You are almost certainly not going to self-host a 1.02T model. That is fine — that is not what this buys you.

What an MIT model buys is a floor under your vendor negotiation and a genuine exit. Every API model you use today comes with terms that can change: price, rate limits, refusal behavior, deprecation dates, whether your sector is allowed. We have written about all five happening this year. An open-weights model under MIT cannot be repriced or retracted. The copy you downloaded is the copy you have.

That matters in three concrete places:

Regulated or offline work. If a workload cannot leave your network, a hosted API is not an option at any price. An MIT model with a 1M context is.

Benchmark leverage. Knowing a free model reaches 80% of what you pay for changes what you are willing to pay for the last 20%. Run your own comparison before your next renewal.

Continuity. Download the weights of a model that does your job acceptably and store them. You will probably never serve them. You are buying the option, and the option costs disk space.

The trap is assuming open weights means free. Serving a 15B-active MoE at production latency is real GPU spend and real ops work. The right posture for most small teams is an API by default with a tested open-weights fallback — not a migration.

Sources: MiMo-V2.6-Pro-RL model card, MiMo-V2.6-Flash-RL model card.