22 September 2026
Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash on 21 September 2026, two open-weight models published under an MIT licence. MiMo-V2.6-Pro costs $0.435 per million input tokens and $0.87 per million output tokens on Xiaomi’s own API, and scores 46.32 on the Artificial Analysis Intelligence Index v4.3. That is level with Grok 4.7, which SpaceXAI released the same day at roughly five times the token price.
What are MiMo-V2.6-Pro and MiMo-V2.6-Flash?
Both models are omnimodal, which means they take text, images, audio and video in and write text out. Both use a sparse mixture-of-experts design, which switches on only part of the model for each token. That is why the running cost tracks the active parameters rather than the full size.
Here is what Xiaomi has published:
- MiMo-V2.6-Pro: 1.02 trillion total parameters, 42 billion active per token
- MiMo-V2.6-Flash: 309 billion total parameters, 15 billion active per token
- Context window: 1,048,576 tokens on both models
- Licence: MIT, so commercial use and self-hosting are allowed
- Weights: free to download from Hugging Face
- Hosted access: Xiaomi AI Studio, MiMo Code, MiMo Desktop, the MiMo API platform and OpenRouter
A third variant, MiMo-V2.6-Pro-UltraSpeed, runs at about twenty times the output speed of Pro for ten times the price. Xiaomi also trained this generation in public, livestreaming the reinforcement learning runs. It puts the final stage at 30 large steps over roughly 750,000 trajectories in under six days, at $2.62 million for Pro and $850,000 for Flash. Those are Xiaomi’s training costs, not yours.
MiMo-V2.6 benchmarks and price against Grok 4.7 and Gemini 3.8 Flash
Read the table for the gap between what MiMo-V2.6-Pro scores and what it charges.
| What is measured | MiMo-V2.6-Pro | Named alternative |
|---|---|---|
| Artificial Analysis Intelligence Index v4.3 (broad reasoning) | 46.32 | 46 (Grok 4.7), 44 (Grok 4.6), 41 (Gemini 3.8 Flash) |
| DeepSWE v1.1 (agentic bug fixing) | 71.9 | 71.0 (Grok 4.7) |
| Terminal Bench 2.1 (command line tasks) | 89.9 | no rival figure published alongside it |
| AutomationBench (computer use) | 53.1 | no rival figure published alongside it |
| Price per million input tokens | $0.435 | $2.00 (Grok 4.7) |
| Price per million output tokens | $0.87 | $6.00 (Grok 4.7) |
The Intelligence Index is a third-party leaderboard. The agent scores are Xiaomi’s own evaluations, and Xiaomi does not name the harness behind them, so treat DeepSWE v1.1, Terminal Bench 2.1 and AutomationBench as vendor-run until an outside team reproduces them. MiMo-V2.6-Flash sits at $0.14 and $0.28 per million tokens, and cached input on Pro drops to $0.004 per million.
What Xiaomi is not saying about MiMo-V2.6
An MIT licence on a 1.02 trillion parameter model is a smaller gift than it sounds. You still have to hold every parameter in memory to serve the model, even though only 42 billion of them fire per token. Xiaomi publishes no hardware guidance and no throughput figures, so the cost of self-hosting Pro is the one number a team weighing weights against an API cannot work out from the announcement.
Claude Opus 5 is still ahead on DeepSWE v1.1 and on Terminal Bench 4.0, and Xiaomi does not dispute that. What it does is lead with the Intelligence Index tie against Grok 4.7. That index is a broad average, and an average hides the agent benchmarks where the gap to Anthropic is real and where most people actually want to use a model like this.
What this means
Worth testing now, and the version to test is Flash rather than Pro. At $0.14 and $0.28 per million tokens with 15 billion active parameters, MiMo-V2.6-Flash is the model a four-person team running code review or document triage agents can both afford on the API and realistically serve on its own hardware. Pro is the one that wins the leaderboard, but 1.02 trillion parameters of MIT-licensed weights is a research asset for most companies and an inference bill for the rest.
The wider point is about price floors. Xiaomi has matched Grok 4.7 on a general reasoning index at about a fifth of the token price, with the weights attached. That is not a claim on the frontier, where Claude Opus 5 still leads the agent benchmarks. It is a claim on the middle of the market, where most production work sits, and it puts pressure on anyone charging $2 per million input tokens for work that does not need a frontier model. If you already pay a hosted vendor for high volumes of routine work, run your own evaluation set through Flash this week. If your workload lives or dies on agent benchmarks, stay where you are.
For more information, visit the official announcement of MiMo-V2.6 on the Xiaomi MiMo site.
