Xiaomi released MiMo-V2.6 as three MIT-licensed checkpoints, and OpenRouter lists the Pro model as released September 21, 2026. The MiMo-V2.6-Pro-RL model card on Hugging Face describes a sparse MoE with 1.02T total and 42B activated parameters. Artificial Analysis scores Pro 46, and the API costs $0.435 per 1M input tokens and $0.87 per 1M output. Self-host only when your API bill tops the cost of an 8- to 16-GPU cluster.
What happened: three MIT checkpoints, one independent score
Xiaomi published Pro-RL, Flash-RL and a 9B distill, each with "License: mit" on its model card. Xiaomi's public RL dashboard shows a total post-training cost counter of $3,474,715 for runs that started September 15. The Pro-RL card links a technical report PDF.
What did the public RL livestream show?
The dashboard shows the Pro run started September 15, 2026 at 10:32 UTC and stopped at step 30 after 5d 07:29:21. The Flash run started at 15:16 UTC and stopped September 19 at step 30 after 3d 11:05:40. Its notices feed admits the failures: "the pro run restarted at step 17 due to a GPU OOM issue caused by expert load imbalance." Xiaomi also removed the cyber dataset from the next Pro run "since we observed some bad patterns in the rollout logs," and restarted Flash "from step 15." The counter covers RL post-training only, not pretraining, so $3.47M is not the cost of building the model. We found no third-party audit of the total.
- What changed: MiMo-V2.6 weights are open under MIT, and the V2.5 API models are retired.
- Who's affected: developers choosing between self-hosting and calling the API.
- What to do: move V2.5 model IDs to V2.6, then run the break-even rule below.
Which MiMo-V2.6 checkpoint fits your hardware?
Pro and Flash share a 1M-token context, while the 9B model is a fine-tune of Qwen3.5-9B on MiMo-generated data. Weight sizes below are our BF16 estimates (parameters x 2 bytes), not vendor figures.
| Checkpoint | Parameters | Context | License | BF16 weights (estimate) |
|---|---|---|---|---|
| MiMo-V2.6-Pro | 1.02T total / 42B activated | 1M tokens | MIT | about 2.0TB |
| MiMo-V2.6-Flash | 309B total / 15B activated | 1M tokens | MIT | about 0.62TB |
| MiMo-V2.6-Distill-Qwen-9B | 9B | not stated | MIT | about 18GB |
Is the 46 independent, and what is vendor-reported?
The 46 is independent: the Artificial Analysis page for MiMo-V2.6-Pro scores it on Index v4.3.2, against a median of 19. That places it 27th of 706 models overall. The Batch, a press outlet, reports it ranks first among open-weights models.
AA also lists a speed of about 50 tokens per second, "slower than average (71)". It puts the cost of running its index at $0.13 per task.
Xiaomi's pricing page lists a hosted-only MiMo-V2.6-Pro-UltraSpeed tier it says delivers "output up to 20x faster" (Xiaomi's claim), plus a Batch API for mimo-v2.6-flash and mimo-v2.6-pro. Neither has open weights.
Most other numbers come from Xiaomi's own table. Independent checks exist for one row.

| Benchmark | Pro (Xiaomi) | Flash (Xiaomi) | Independent |
|---|---|---|---|
| Terminal Bench 4.0 | 34.9 | 28.8 | AA 34.8, Vals 31.3 |
| Terminal Bench 2.1 | 89.9 (Claude Opus 5: 89.1) | 87.6 | none |
| CyberGym | 94.0 | 95.1 | none |
| ProgramBench | 26.5 (Opus 5: 37.0, GPT-5.6 Sol: 25.0) | 26.0 | Vals 0.5% (different harness; not comparable, not a failure) |
| ExploitBench | 47.9 (Claude Opus 5: 70.0, GPT-5.6 Sol: 78.5) | 25.3 | none |
| DeepSWE v1.1 | 71.9 (card), 72.57 (RL dashboard) | 67.9 (card), 65.68 (dashboard) | none |
The ExploitBench and Terminal Bench 4.0 gaps to closed models come from Xiaomi's own card. On Terminal Bench 4.0, the card lists Claude Opus 5 at 49.0 against Pro's 34.9.
Should you self-host Pro or call the API?
Call the API unless your monthly bill exceeds your GPU count (8 with vLLM, 16 with the SGLang recipe) x your hourly GPU rate x 730 hours. The SGLang recipe is --tp 16 --dp 2 --ep 16 --nnodes 2, with --reasoning-parser mimo --tool-call-parser mimo. The card's vLLM recipe uses --tensor-parallel-size 8, so check whether one 8-GPU node fits your context and throughput before pricing two.

- Price your API use. On OpenRouter, Pro lists at $0.435 input and $0.87 output per 1M tokens. DeepInfra charges $0.387 / $0.783. Flash starts at $0.10 / $0.24.
- Price the cluster. Multiply your GPU count (8 or 16) by your hourly rate and 730 hours.
- Check quantized options. The Pro-RL page shows an fp8 tag and nine quantized models in its model tree. Whether the fp8 files are official is unconfirmed.
- Use the card's sampling. Set temperature 1.0 and top_p 0.95.
If you mix models by task, our guide to AI model routers and their real cost math covers routing across providers.
What to do now
- October 21, 2026, 10:00 (UTC+8): Xiaomi retired MiMo-V2.5-Pro and MiMo-V2.5, and "their model IDs no longer work." Switch to the V2.6 series.
- Before a long contract: compare Pro with Kimi K3 and Qwen3.8 Max on your own prompts.
What's still unconfirmed
- CNY API prices. Xiaomi's pricing page showed no price table when we checked. Check the vendor page.
- RL framework. The Batch reports 7,000+ RL environments and training code; we could not find them on a Xiaomi page.
- Per-run costs. The dashboard shows only the $3,474,715 total. The Batch reports $2.6M for Pro and $0.9M for Flash (press).
- Distillation allegation. The Batch reports that Anthropic alleges Xiaomi routed user sessions to Claude for training data; Xiaomi had not responded.




