Xiaomi’s new MiMo models carry an unverified top-6 claim

On September 21, Xiaomi released its open-weight MiMo-V2.6-Pro and MiMo-V2.6-Flash, and said how much training them cost. Testing Catalog reports that the reinforcement-learning run behind each model took under six days and cost about $2.62 million for Pro and $850,000 for Flash.

Xiaomi’s site and OpenRouter confirm both models have shipped. A post on Threads said Pro had reached number 6 on the Artificial Analysis Intelligence Index and was the top open-weight model there. We could not verify that rank. The post it traces to is dated September 13, 2025, so it may be a snapshot from more than a year ago and not the current board. Xiaomi’s own announcement, linked from its release page, says: “Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks.” Alongside the weights, Xiaomi says it is publishing the technical report, the training environments and the RL code.

Thirty steps in six days

The release is built around scaling reinforcement learning on complex tasks with verifiable outcomes, where a program can check whether the model actually finished the job. Each model completed 30 RL steps across roughly 750,000 trajectories. A trajectory is one full attempt at a task, from first action to scored result. Each update used 1,568 samples, with contexts of up to one million tokens, and the tasks covered coding, general-agent work, visual problems and cybersecurity. Xiaomi calls both models natively omnimodal and aims them at coding, visual tasks and computer use.

Xiaomi froze the router to keep the run from drifting. The router is the part of a mixture-of-experts model that picks which experts process each token. Against reward hacking, where a model learns to satisfy the checker without solving the task, it used adversarial evaluation, anomaly detection and cross-checks between verifiers.

The gain Xiaomi reports is on DeepSWE v1.1, an agentic coding benchmark. Over the run, Flash went from 48.8 to 65.68 and Pro from 58.4 to 72.57. In August, Alibaba’s Qwen3.8 Flash scored 58.7 on the same test. That is just above Pro’s starting point and about seven points below where Flash ended up after RL.

The ranking and the Opus comparison

Testing Catalog puts Pro at 46.32 on version 4.3 of the Artificial Analysis index. Xiaomi says that puts it ahead of Kimi K3 and Qwen3.8 Max, making it the highest-scoring open-source model in its comparison. The parity with Claude Opus 5 and GPT-5.6 Sol is Xiaomi’s claim, and its own wording covers “most” agent benchmarks, not all of them. We could not verify that comparison independently.

What it costs

API prices stay at V2.5 levels. Flash costs $0.14 per million uncached input tokens and $0.28 per million output tokens. Pro costs $0.435 and $0.87. Pro-UltraSpeed is a separate serving tier of Pro, not new weights. It promises up to 20 times faster output at the same quality for latency-sensitive work, and it costs ten times more.

Pro and Flash are available now in AI Studio, MiMo Code, MiMo Desktop, Xiaomi’s MiMo API Platform and OpenRouter. MiMo Desktop is leaving early access with both models included, and UltraSpeed is offered for real-time workflows.

“Vibe World”

Xiaomi is also pitching the models for work beyond conventional software, under the label “Vibe World.” From an image, a video or a text prompt, MiMo-V2.6 can coordinate agents that build interactive 3D scenes and then test them visually. Xiaomi says the models can also create Blender assets, control a Franka Panda robotic arm from camera feeds, produce frontends and presentations, assemble videos, and write music as scores and MIDI.

Xiaomi also showed research demos. One screened materials for capturing PFAS chemicals. In another, the model helped formalize a Lean 4 theorem in more than 6,000 lines of kernel-verified code.