Xiaomi · 2026-09-21 · major
MiMo-V2.6 — Xiaomi's trillion-parameter model ships under MIT
MiMo-V2.6 is Xiaomi's new model series: a 1.02T-parameter Pro, a cheaper Flash, and a 20x-faster UltraSpeed variant. Weights for Pro and Flash are on Hugging Face under the MIT license with a 1M-token context.
Xiaomi released MiMo-V2.6 with open MIT weights, a 1M-token context, and cybersecurity scores far above its own V2.5 line.
Quick facts
| Maker | Xiaomi |
|---|---|
| Parameters (Pro) | 1.02T total, 42B active |
| Context window | 1M tokens |
| License | MIT |
| Variants | Pro, Flash, Pro-UltraSpeed, Distill-Qwen-9B |
| Weights | Hugging Face (XiaomiMiMo) |
| Serving | SGLang or vLLM |
Benchmarks
| MiMo-V2.6 Pro | 71.9% | |
|---|---|---|
| MiMo-V2.6 Flash | 67.9% | |
| Claude Opus 5 | 74% | |
| GPT-5.6 Sol | 73% | |
| Claude Fable 5 | 70% |
| MiMo-V2.6 Pro | 76.9% | |
|---|---|---|
| MiMo-V2.6 Flash | 73.6% | |
| Claude Opus 5 | 80.6% | |
| GPT-5.6 Sol | 74.9% | |
| Claude Fable 5 | 77.9% |
| MiMo-V2.6 Pro | 89.9% | |
|---|---|---|
| MiMo-V2.6 Flash | 87.6% | |
| Claude Opus 5 | 89.1% | |
| GPT-5.6 Sol | 88.8% | |
| Claude Fable 5 | 84.3% |
What is it?
Three new models join Xiaomi's MiMo line as MiMo-V2.6: a trillion-parameter Pro built for long-horizon and cybersecurity work, a cheaper full-modality Flash, and an UltraSpeed build of Pro that Xiaomi says runs up to 20x faster for latency-sensitive jobs. Weights for MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL, and MiMo-V2.6-Distill-Qwen-9B are published on Hugging Face under the MIT license.
How does it work?
The Pro model is a sparse mixture of experts: 1.02 trillion total parameters with 42 billion active per token, 384 routed experts of which 8 fire per token, and a 1M-token context. Xiaomi credits the jump to scaled reinforcement learning — fully asynchronous GRPO over batches of 1,568 prompts by 16 rollouts per step, plus groupwise agentic grading and multi-prefix multi-teacher on-policy distillation. A 681M-parameter MiMo ViT reads images and a 308M AudioTokenizer handles sound.
Why does it matter?
An MIT license on a trillion-parameter omni-modal model is unusual — teams can self-host, fine-tune, and ship MiMo-V2.6 commercially with no usage restrictions. On Xiaomi's own table Pro reaches 89.9% on Terminal Bench 2.1, above Claude Opus 5 at 89.1%, and 94.0% on CyberGym against 40.0% for MiMo-V2.5 Pro. Xiaomi recommends serving it with SGLang or vLLM.
Who is it for?
teams self-hosting open-weight frontier models
Frequently asked questions
- Is MiMo-V2.6 open source?
- The weights are. Xiaomi publishes MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL, and MiMo-V2.6-Distill-Qwen-9B on Hugging Face under the MIT license, which places no restriction on commercial use or modification. The hosted API variants — mimo-v2.6-pro, mimo-v2.6-flash, and mimo-v2.6-pro-ultraspeed — are served by Xiaomi separately through its own platform.
- What hardware do you need to run MiMo-V2.6 Pro?
- MiMo-V2.6-Pro-RL is a 1.02 trillion parameter mixture of experts with 42 billion parameters active per token, so it needs a multi-GPU node rather than a single card. Xiaomi recommends deploying it with SGLang or vLLM using specific tensor-parallel configurations. The 9B Distill-Qwen build exists for people who cannot host the full model.
- How does MiMo-V2.6 compare to MiMo-V2.5?
- The gap is large on agentic work. On Xiaomi's published table MiMo-V2.6 Pro scores 71.9% on DeepSWE v1.1 against 19.0% for MiMo-V2.5 Pro, 76.9% against 49.1% on Toolathlon-Verified, 89.9% against 65.2% on Terminal Bench 2.1, and 94.0% against 40.0% on CyberGym.
- What is MiMo-V2.6-Pro-UltraSpeed for?
- MiMo-V2.6-Pro-UltraSpeed is Xiaomi's latency build: the company describes it as flagship V2.6-Pro performance running up to 20x faster, aimed at real-time and latency-sensitive workloads. Pick it when response time matters more than cost per token. The plain mimo-v2.6-flash variant is the cheaper option for high-frequency calls at large scale.
Try it
Pull XiaomiMiMo/MiMo-V2.6-Pro-RL from Hugging Face and serve it with SGLang or vLLM.