MiMo-V2.6-Pro
Flagshipby Xiaomi
MiMo-V2.6-Pro is the flagship of Xiaomi's MiMo-V2.6 series, released September 21, 2026 with open weights under the MIT license as XiaomiMiMo/MiMo-V2.6-Pro-RL. It is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per token, natively omnimodal across text, image, video, and audio input, with a 1 million token context window and up to 131,072 output tokens. The API rate did not change from V2.5: $0.435 per million input tokens, $0.87 output, and $0.0036 for cached input, which is roughly a tenth of Claude Opus 5.5 on input and a twenty-third on output. The independent numbers are why it matters. Artificial Analysis scores it 46 on its Intelligence Index, the top open-weights result on that board and level with Grok 4.7, and the Vals Index places it at 59.47 percent, ahead of DeepSeek V4.1 Flash. The vendor table is strongest on agent work: 71.9 on DeepSWE v1.1, 89.9 on Terminal-Bench 2.1, and 53.1 on AutomationBench, the last ahead of the Claude Opus 5 row on Xiaomi's own chart. Terminal-Bench 4.0 at 34.9 is where the gap to the closed frontier shows. Xiaomi reports 82.0 on OSWorld-Verified, a different task set from OSWorld 2.0, so it does not appear in that column on this site. The training story is the other reason to read the card: one mixed reinforcement learning run across coding, general agents, visual, and cybersecurity tasks at 1,568 prompts per update, with the technical report, environments, and RL code published alongside the weights. A Pro-UltraSpeed tier serves the same model faster at ten times the price.
Input Price
$0.43
per 1M tokens
Output Price
$0.87
per 1M tokens
Context Window
1.0M
tokens
Released
2026-09
Open source
Capabilities
Key Strengths
- ✓MIT-licensed open weights at 1.02T total and 42B active parameters
- ✓Top open-weights score on the Artificial Analysis Intelligence Index at 46
- ✓$0.435/$0.87 per 1M tokens, unchanged from V2.5
- ✓Native text, image, video, and audio input in a 1M token context
- ✓DeepSWE v1.1 71.9 and Terminal-Bench 2.1 89.9, vendor-reported
- ✓Technical report, training environments, and RL code published
Best For
- ▸Self-hosted agentic coding where license terms matter
- ▸Omnimodal agents that read video and audio in the same run
- ▸Cost-sensitive long-context work at frontier-adjacent quality
- ▸Research on scaled reinforcement learning with published training code
Benchmark Scores
| Benchmark | Score | Description |
|---|---|---|
| Terminal-Bench 4.0 | 24.8 | Long-horizon agentic work in a terminal across software, science, ML, operations, hardware, security, and media (66 tasks, all-or-nothing verifiers) |
Scores sourced from public benchmark datasets. See full benchmark leaderboard for all models.
Pricing Details
Input tokens
$0.43
per 1M tokens
Output tokens
$0.87
per 1M tokens
Estimated cost per 1K requests
$0.87
~1K input + ~500 output tokens avg
Prices are subject to change. Check the official documentation for current pricing. See the cost calculator for detailed estimates.
Open Source Model
MiMo-V2.6-Pro is free to download and self-host under the MIT. Hosted API pricing varies by provider (e.g., Together, Fireworks, Groq). See our open source LLM guide for deployment options.