Skip to content
NewsModel release

Xiaomi MiMo-V2.6-Pro Claims Top Open-Weights Score on Artificial Analysis Index at 46, Ahead of Kimi K3

· by Pondero Newsdesk

The short version

Xiaomi released the MiMo-V2.6 series on September 22, 2026, with the 1T-parameter Pro variant scoring 46 on the AA Intelligence Index to claim the top open-access model position worldwide, under an MIT license.

Xiaomi MiMo-V2.6-Pro Claims Top Open-Weights Score on Artificial Analysis Index at 46, Ahead of Kimi K3

A company best known for budget smartphones and TVs now holds the top open-weights AI ranking. Xiaomi released the MiMo-V2.6 series on September 22, 2026, with the Pro variant reaching 46 on the Artificial Analysis Intelligence Index and surpassing Kimi K3 and Qwen3.8 Max for the top open-access model position worldwide, per Xiaomi's release page.

What

MiMo-V2.6-Pro is a fully multimodal model with 1 trillion parameters. Xiaomi released the weights under an MIT license on Hugging Face alongside a technical report and the reinforcement-learning training code. The series has three variants: Pro for complex and long-horizon tasks, Flash for high-frequency production calls at lower cost, and Pro-UltraSpeed which delivers the same accuracy as Pro at up to 20x the generation speed, per Xiaomi. All three carry a 1 million-token context window.

On DeepSWE v1.1, an out-of-sample software engineering benchmark, Pro scored 72.6 and Flash scored 65.7, up from 58.4 and 48.8 on the V2.5 series, gains of roughly 14 and 17 points respectively, according to Xiaomi's own benchmarks. Xiaomi also states that Pro performance is on par with Claude Opus 5 and GPT-5.6 Sol on multi-agent benchmarks and the Design Arena leaderboard. No independent third-party benchmark results had been published as of the September 22 release date.

Pricing is unchanged from the V2.5 series: Pro costs $0.435 per million input tokens and Flash costs $0.14 per million input tokens on the overseas pay-as-you-go tier, per Xiaomi's pricing page. Xiaomi states RL training cost roughly $850,000 for Flash and $2.62 million for Pro, each completed across 6 days with approximately 750,000 total trajectories.

Operators can self-host the top open model under MIT

The Artificial Analysis index is the same benchmark where StepFun's Step 5 Preview landed at 44 last week. A consumer electronics manufacturer claiming the top open-weights slot changes who operators need to watch in the self-hosting market. MIT licensing means no restrictions on commercial use or redistribution, and operators with GPU infrastructure can run MiMo-V2.6-Pro without per-token API billing: the same dynamic that drove adoption of DeepSeek R2 and Kimi K3.

Flash at $0.14 per million input tokens is priced for high-frequency production use. Pro-UltraSpeed solves a different problem: latency-sensitive workloads where smaller, lower-accuracy models are typically the only option. Together the three variants cover the cost, accuracy, and speed tradeoffs most self-hosting teams face.

One caveat: the DeepSWE scores, the intelligence-index rank, and the comparisons to Claude Opus 5 and GPT-5.6 Sol all come from Xiaomi's own report. Independent evaluations from labs such as LiveCodeBench had not published results as of the release date.

What to watch next

Whether inference providers such as Together AI, Fireworks, or Groq add MiMo-V2.6-Pro hosting is the near-term access question for teams without self-hosting hardware. Independent benchmark results from third parties would confirm or revise the internal DeepSWE scores and intelligence-index position.

Sources