Xiaomi MiMo-V2.6 Pro: What the Top Open Model Offers

The Open-Weight Crown Changes Hands

Xiaomi released MiMo-V2.6 Pro and MiMo-V2.6 Flash on September 21 (September 22 in Beijing), publishing both under the permissive MIT license with weights on Hugging Face. The Pro model scores 46 on the Artificial Analysis Intelligence Index, the highest figure any open-weight model has posted on that leaderboard. For anyone building on models they can download, inspect and host themselves, that is the headline.

• Scores: 46.32 on the Artificial Analysis Intelligence Index for MiMo-V2.6 Pro, ahead of GLM-5.3 (45) and Kimi K3 (44), per TechNode and VentureBeat
• Size: Pro has 1.02 trillion total parameters with 42 billion active per token; Flash has 310 billion total and 15 billion active, according to VentureBeat
• Context: 1 million tokens of input and up to 128K tokens of output, with text, image, video and audio input
• API pricing: Pro costs $0.435 per million input tokens and $0.87 per million output; Flash costs $0.14 and $0.28

How Good Is MiMo-V2.6 Pro, Really?

Good enough to change the conversation. VentureBeat notes that its score of 46 ties SpaceXAI's Grok 4.7, the closed model we covered in Grok 4.7's GitHub Copilot debut just yesterday, and sits above Gemini 3.8 Flash at 41. TechNode puts the leading closed models at 53, so a gap to the very top remains. But for the first time, an open-weight model is trading blows with a brand-new frontier release on an independent composite index.

On coding, Xiaomi reports big jumps on DeepSWE v1.1: Flash climbed from 48.8 to 65.68 and Pro from 58.4 to 72.57, according to FoneArena. Those are Xiaomi's own numbers and deserve independent replication, but they line up with the direction of the Artificial Analysis result.

What Did Xiaomi Change in Training?

The story here is reinforcement learning at scale, done efficiently. According to TechNode and FoneArena, each model completed 30 RL steps in under six days using roughly 750,000 trajectories, processing about 3.5 to 3.7 billion tokens per step. Xiaomi puts the RL cost at about $2.62 million for Pro and $850,000 for Flash.

Those figures matter because they are small. Frontier-level post-training has often been described as something only the largest labs could afford. A multi-million-dollar RL run that lifts a mixture-of-experts model to the top of the open-weight field suggests the technique is becoming reproducible, and according to FoneArena, Xiaomi has published the technical report, training environments and RL code to help others do exactly that.

Fuli Luo, a former DeepSeek researcher now at Xiaomi, framed it plainly in comments reported by VentureBeat: the team chose to dedicate several dozen people to one goal over an extended period, which was scaling up RL.

What Can Developers Do With It Today?

Quite a lot. The MiMo-V2.6 Pro API speaks both OpenAI and Anthropic protocols, so most existing agent tooling can point at it with a base URL change. According to FoneArena, the models are live across Xiaomi AI Studio, MiMo Code, the MiMo API Platform and OpenRouter, and the MiMo Desktop client has moved to an official release. There is also a faster Pro-UltraSpeed tier at $4.35 input and $8.70 output per million tokens for latency-sensitive work.

Self-hosting the full Pro model is a data-center job at 1.02 trillion parameters, much like the Kimi K3 open-weight release this summer. Flash, with 15 billion active parameters, is the more realistic target for well-equipped workstations once community quantizations arrive. Xiaomi also shipped a MiMo-V2.6-Distill-Qwen-9B model, per TechNode, which is small enough for a single consumer GPU.

Why This Release Matters

Open-weight models are how most of the world will actually run AI: on their own servers, under their own rules, at their own cost. MiMo-V2.6 shows that the gap between what you can download and what you can only rent keeps shrinking, and that careful RL can close more of it than raw scale alone. Follow the rest of this week's model news in our AI coverage.

Sources: VentureBeat — September 21, 2026; TechNode — September 22, 2026; FoneArena — September 22, 2026; Xiaomi MiMo — September 2026.