Xiaomi open-sources MiMo-V2.6 multimodal AI models

MiMo-V2.6 is Xiaomi's newly open-sourced multimodal AI series, bringing native 3D spatial reasoning, multimodal perception, and direct computer-use capabilities to open weights.

The lineup includes the flagship native multimodal MiMo-V2.6-Pro, the faster MiMo-V2.6-Flash, a distilled MiMo-V2.6-Distill-Qwen-9B model, and an UltraSpeed variant delivering up to 20x faster output. Xiaomi also released a dedicated desktop client alongside the weights.

What it does

The marquee upgrade in MiMo-V2.6 is spatial reasoning and autonomous computer interaction. The models can interpret 3D visual space and execute computer-use tasks directly through mouse and keyboard actions on the desktop.

Benchmark performance is notably strong. On the Artificial Analysis Intelligence Index, MiMo-V2.6-Pro scored 46 (specifically 46.32), edging out Kimi K3 at 44 and GLM-5.3 at 45. That makes it the highest-ranked open-weight model on the board. In coding benchmarks, it also boosted DeepSWE v1.1 scores by roughly 14 points.

Xiaomi is realistic about the ceiling, though. The company openly noted that top-tier closed-source models still lead the index at 53 points.

How it was built

The secret sauce here is rapid, large-scale reinforcement learning.

Xiaomi pushed through 30 reinforcement-learning steps across roughly 750,000 trajectories in less than six days. The team was refreshingly transparent about compute bills: the RL stage alone cost $2.62 million for the Pro model and $850,000 for the Flash version.

Alongside the open weights, Xiaomi released research resources detailing the reinforcement learning setup. For teams preferring cloud endpoints, API pricing remains unchanged from MiMo-V2.5.

Why it matters

If you are an engineer building local computer-use agents, desktop automation, or spatial reasoning pipelines, MiMo-V2.6-Pro is an immediate download that matches or beats existing open alternatives.

More importantly, putting a public $2.62 million price tag on the reinforcement learning phase demystifies what it actually costs to push open weights to the top of the leaderboard. Xiaomi proved that open multimodal models can keep closing the gap with closed giants, provided someone is willing to foot the compute bill.