Quick Facts
- MiMo-V2.6-Pro scores 46 on Artificial Analysis' Intelligence Index, topping all open weights models and beating xAI's Grok 4.6 and Google's Gemini 3.8 Flash
- The model costs $0.435 per million input tokens, compared to $5.00 for Claude Opus 5, which sits just five index points higher
- Xiaomi livestreamed the entire reinforcement learning post-training run publicly, reporting a combined cost of $3.47 million for both new models
Xiaomi has released MiMo-V2.6-Pro, and it now holds the top spot among open weights models worldwide. The model scored 46 on Artificial Analysis' Intelligence Index, a 20-point jump from its predecessor MiMo-V2.5-Pro, which scored 26.
The leap puts MiMo-V2.6-Pro ahead of proprietary models including xAI's Grok 4.7, which also debuted the same day at 46, xAI's Grok 4.6 at 44, and Google's Gemini 3.8 Flash at 41. It also surpasses fellow Chinese open weights models DeepSeek V4.1 Flash (39) and DeepSeek V4.1 Pro (36).
Before this release, the open weights crown was shared by Z.AI's GLM-5.3 and Moonshot's Kimi K3, both at 44. The title had traded between Chinese labs for months.
Architecture and Capabilities
MiMo-V2.6-Pro is a 1.02-trillion-parameter mixture-of-experts model with approximately 42 billion parameters active per token. The model supports a 1-million-token context window and native multimodal input.
Xiaomi also released MiMo-V2.6-Flash, a smaller sibling built on the same architecture with 309 billion total parameters and 15 billion active. Flash targets high-volume production workloads while retaining the 1-million-token context window and multimodal support.
On the coding benchmark DeepSWE v1.1, MiMo-V2.6-Pro scored 71.9 against 67.9 for Flash. Both trail Claude Opus 5 and GPT-6 Astra at 74.0 on that specific test, though Xiaomi says the Pro model performs on par with those proprietary systems across most agent benchmarks.
Pricing That Changes the Conversation
MiMo-V2.6-Pro costs $0.435 per million uncached input tokens and $0.87 per million output tokens. MiMo-V2.6-Flash comes in at $0.14 per million input tokens and $0.28 per million output tokens.
For comparison, Anthropic's Claude Opus 5, which scores five index points higher at 51, costs $5.00 per million input tokens and $25.00 per million output tokens. Artificial Analysis pegged MiMo-V2.6-Pro's cost at $0.13 per Intelligence Index task, with the full evaluation run costing $206.66 total.
For software teams currently paying top-tier prices for proprietary models, those numbers are hard to ignore. A high-scoring open weights model at this price point gives buyers real negotiating leverage with incumbent providers.
The $3.47 Million Public Training Run
One of the most unusual aspects of this launch was transparency. Since Sept. 15, 2026, Xiaomi has been streaming a live training dashboard at mimo.xiaomi.com/rl/, showing steps, reward curves, token throughput, benchmark scores, and a running cost counter in real time.
Fuli Luo, the MiMo team lead and a former DeepSeek researcher, announced the stream on X, calling it the first public livestream of a full frontier-scale reinforcement learning post-training run. Xiaomi reports spending $2.62 million on the Pro run and $854,044 on Flash, totaling roughly $3.47 million across both.
Each run completed 30 reinforcement learning steps over roughly 750,000 trajectories in under six days. Xiaomi describes its training approach as "You Only RL Once," mixing programming, general agent, vision manipulation, and cybersecurity tasks into a single reinforcement learning batch.
Multimodal and Robotics Reach
The new models add 3D spatial reasoning and computer-use capabilities. Xiaomi calls this expansion "Vibe World," describing workflows where a user inputs images, video, or text and the model breaks the request into tasks completed through multi-agent collaboration, including 3D scene construction and interactive logic programming.
On the robotics side, Xiaomi says MiMo-V2.6 can take multi-view camera images as input and control a Franka Panda robotic arm through a visual feedback loop to complete object grasping and color matching tasks. The company frames this capability as part of a longer push toward recursive self-improvement, describing it as "a key step in our exploration of the RSI path."
The open weights release means any team can download and deploy the model without paying per-token API fees, a factor that will push enterprise buyers to reconsider vendor commitments made before this week.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.