Xiaomi's MiMo-V2.6 tops open-source ranking after six-day, $3.5M live RL run

The Pro model has 1.02 trillion parameters and 42 billion activated parameters. It scored 46 on the Artificial Analysis Intelligence Index, ranking first among open-source models. Hugging Face CEO praised it, saying 'Awesome.' In many Agent evaluations, QbitAI reported, Pro has approached Claude Opus 5 and GPT-5.6 Sol.

Pro ran for five days and three hours at a cost of about $2.6 million, while Flash ran for three days and ten hours at about $900,000. Each step contained 1,568 prompts, with 16 routes per prompt, generating more than 25,000 rollouts per round. Pro processed roughly 81 billion to 111 billion training tokens, equivalent to filling more than 80,000 million-token context windows.

After 30 steps, Flash's average pass rate improved 25% relative, and Pro's improved 12% relative. On DeepSWE v1.1, an out-of-sample test for coding agents working in real open-source repositories, Pro rose from 58.4 to 72.57, and Flash rose from 48.7 to 65.68. On AutomationBench, Pro scored 53.1, above Claude Opus 5's 50.3 and GPT-5.6 Sol's 45.8.

Xiaomi kept MiMo-V2.6 API pricing at V2.5 levels: Pro at about 3 yuan per million input tokens and 6 yuan per million output tokens, and Flash at about 1 yuan and 2 yuan. Artificial Analysis estimated that MiMo-V2.6-Pro costs an average of $0.13 to complete one Intelligence Index task. That compares with $3.74 for the xHigh version of Grok 4.7, which also scored 46 but cost about 29 times more, according to QbitAI. The MiMo-V2.6-UltraSpeed mode reaches up to about 900 TPS.

Xiaomi released the Pro and Flash model weights, a full technical report, more than 7,000 RL task environments, an end-to-end RL training framework, and combinable mini-harnesses. It also released MiMo-V2.6-Distill-Qwen-9B, an SFT model built on Qwen3.5-9B and trained on 77.4 billion tokens of MiMo-generated data. The company said the release is intended to give the community a stronger starting point for Agent RL.

The open release drew rapid reactions. Ai2 post-training lead Nathan Lambert said people who understand open-source models already knew Xiaomi was preparing something with MiMo, and praised the combination of a top-scoring open-source model and a public RL dashboard. Hugging Face former researcher Elie Bakouch said Xiaomi delivered the model, technical report, more than 7,000 RL environments, and a complete training framework less than a week after starting final RL training. AI researcher Himanshu Raj said the gap between open-source models and the closed-source frontier had been narrowed sharply, and that this time the player was not the familiar Chinese company.

Before the RL run, Xiaomi also tested MiMo-V2.6-Pro on a materials science problem without task-specific training. The model designed new metal-organic frameworks, or MOFs, to capture PFAS, a class of persistent pollutants. It searched literature and patents, formed hypotheses, checked novelty, built simulation environments, and calculated binding strength. It produced two candidate structures, A50 and B50, both UiO-67 zirconium-based frameworks. Simulations showed their PFAS adsorption performance was one million to ten million times that of control material C50. Dou Jinhu, a distinguished researcher at Peking University's School of Materials Science and Engineering, said the model's performance was equivalent to a well-trained doctoral researcher. Xiaomi said the approach has been used in internal new-material R&D, cutting a one-month cycle to two or three days.

Xiaomi also presented MiMo-V2.6 as a multimodal model that can handle code, images, video, 3D, and music. It built a MiMo exhibition hall in which images, video, sound, and 3D models were generated by MiMo, and demonstrated control of a Franka Panda robotic arm through visual feedback for object grasping and color sorting. The company also launched MiMo Desktop, a desktop client that supports multiple agents working together. In one demonstration, a main agent split construction tasks for a 3D Chinese-style voxel world among sub-agents, which ran in parallel while the main agent built the core engine and page structure.

The technical report described an asynchronous training architecture that allows generation, execution, scoring, and training to proceed at the same time. Xiaomi mixed code, general Agent, vision, and cybersecurity tasks in one RL run, and paired task types with different agent harnesses so the model would learn transferable abilities rather than memorize one workflow. QbitAI reported that the team also had to address risks such as models copying answers, GPU memory overflows, scorer failures, and MoE expert load collapse.

Editor's Summary Xiaomi's six-day live-streamed reinforcement learning run for MiMo-V2.6 cost about $3.5 million and produced a 1.02-trillion-parameter open-source model that ranked first on the Artificial Analysis Intelligence Index. Xiaomi released model weights, more than 7,000 RL environments, and a training framework while keeping API prices unchanged. The company also demonstrated the model's use in materials discovery, multimodal generation, and multi-agent desktop tasks.