MiMo-V2.6
Xiaomi's MIT-licensed open-weight pair: MiMo-V2.6-Pro, the top open-weight model on the Artificial Analysis Intelligence Index v4.3.2, and MiMo-V2.6-Flash, a cost-efficient generalist that fits a two-node Faraday cluster.
Model Story
Xiaomi released MiMo-V2.6 on September 21–22, 2026, under the MIT license. Xiaomi also published its training code and reinforcement learning environments, alongside a 9B distilled model. The release comes in two main sizes: MiMo-V2.6-Pro and MiMo-V2.6-Flash.
MiMo-V2.6-Pro is a sparse Mixture-of-Experts model with 1.02 trillion total parameters, 42 billion active, and a 1 million token context window. Artificial Analysis lists it as the #1 open-weight model on its Intelligence Index v4.3.2, with a score of 46 at roughly $0.13 per Intelligence Index task in Artificial Analysis's own harness. That puts it about 7 points behind the GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon tier (53), which costs $1.99–$7.63 per task, and 12 points behind Claude Opus 5.5 at max effort (58).
MiMo-V2.6-Flash is the smaller sibling: 309 billion total parameters with 15 billion active, also MIT-licensed. Artificial Analysis scores it at 38 on the v4.3 index. Artificial Analysis also describes Flash as slower than average and very verbose, which is worth knowing before you choose it for latency-sensitive work.
Faraday Machines ships MiMo-V2.6-Flash as a cost-efficient generalist on the Growth tier. MiMo-V2.6-Pro is available on larger clusters for organizations that want the strongest open-weight model available.
Key Specifications
| Developer | Xiaomi |
|---|---|
| Release Date | September 21–22, 2026 |
| Architecture | Sparse Mixture-of-Experts (MoE) |
| Total Parameters | 1.02 trillion (Pro) / 309 billion (Flash) |
| Active Parameters | 42 billion (Pro) / 15 billion (Flash) |
| Context Window | 1,000,000 tokens (Pro) |
| License | MIT (both models), with open training code and RL environments |
| Size at 4-bit | Roughly 560 GB (Pro) / roughly 170 GB (Flash) |
| Faraday Tier | Flash: Growth (2× Mac Studio, 256GB); Pro: larger clusters of ~6+ units |
| Also Released | 9B distilled model |
Pricing
Artificial Analysis measures MiMo-V2.6-Pro at roughly $0.13 per Intelligence Index task, a small fraction of the $1.99–$7.63 per task it measures for the closed models that score 53. The MIT license means no licensing fees and no usage caps, and on Faraday hardware there are no per-token costs; you pay only for hardware amortization and electricity.
Benchmarks
All scores above are independent Artificial Analysis results on the v4.3 index, which was rescaled in September 2026 with Terminal-Bench 4.0 and AutomationBench-AA. Every model scored lower than on the earlier v4.1 index, so these numbers should not be compared with v4.1 scores elsewhere on this site. On v4.3 the gap between the best open model and the very top closed model widened, while the cost advantage of open models grew.
On-Premises Deployment
At 4-bit quantization MiMo-V2.6-Flash needs roughly 170 GB of memory, which fits on Faraday's Growth tier: two Mac Studio M5 Max units with 256GB of unified memory in total. With only 15 billion active parameters per token, it is a cost-efficient generalist for drafting, analysis and everyday business tasks. Its verbosity matters less on owned hardware, where extra output tokens cost time rather than per-token fees.
MiMo-V2.6-Pro needs roughly 560 GB at 4-bit quantization, which calls for a cluster of about six or more units. Faraday's calculator configures clusters from 4 to 64 units, so organizations that want the top open-weight model on the Artificial Analysis index can run it entirely on hardware they own.
Both models are MIT-licensed, so legal teams can sign off without the review cycles required for custom licenses like those on Kimi K3 or GLM-5.3.
Frequently Asked Questions
Can I run MiMo-V2.6 on-premises?
Yes. MiMo-V2.6-Flash needs roughly 170 GB at 4-bit quantization and runs on Faraday Machines' Growth tier: two Mac Studio M5 Max units with 256GB of unified memory in total. MiMo-V2.6-Pro needs roughly 560 GB at 4-bit, so it calls for a cluster of about six or more units, which Faraday Machines configures on request.
Is MiMo-V2.6 open source?
Yes. Xiaomi released both MiMo-V2.6-Pro and MiMo-V2.6-Flash under the MIT license on September 21–22, 2026, and also published its training code, reinforcement learning environments and a 9B distilled model. The MIT license means no licensing fees or usage caps, and legal teams can approve it without the review cycles that custom licenses such as Kimi K3's or GLM-5.3's require.
How does MiMo-V2.6-Pro compare to GPT-6 Astra and Claude Opus 5.5?
MiMo-V2.6-Pro trails both but costs far less per task. Artificial Analysis lists it as the #1 open-weight model on its Intelligence Index v4.3.2, scoring 46 at roughly $0.13 per task. That is about 7 points behind GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon (53), which cost $1.99–$7.63 per task, and 12 points behind Claude Opus 5.5 at max effort (58).
Is MiMo-V2.6-Flash fast?
Not especially. Artificial Analysis describes MiMo-V2.6-Flash as slower than average and very verbose, and scores it at 38 on its Intelligence Index v4.3. With 15 billion active parameters it suits drafting, analysis and everyday business tasks better than latency-sensitive work. On Faraday Machines hardware its verbosity costs time rather than per-token fees.
Explore Other Models
See how every model compares on price, benchmarks and hardware fit in the AI Model Comparison.