MiMo V2.6 Pro

Latent Space's AINews reports that MiMo-V2.6-Pro debuted as the top open-weights model on the Artificial Analysis Intelligence Index at 46 points, with Xiaomi's own statement describing the series as "natively omnimodal" and MiMo-V2.6-Pro as Xiaomi's most capable model to date. The newsletter highlights an accompanying The writeup contextualizes Xiaomi's emergence as a Chinese frontier lab outside the traditional "six AI Tigers," attributing MiMo-V2.6-Pro's success to transparent live-publishing of final RL training runs by former DeepSeek engineer Fuli Luo. Technical details include large-batch asynchronous training (1,568 samples p

BenchLM lists MiMo-V2.6-Pro with pricing of $0.43 input / $0.87 output per million tokens, $0.004 cached, and a blended price of $0.65, noting speed at 76 tok/s with a 28.79s first-token latency. The model supports a 1M token context window. Coverage is complete for Agentic (9/9 verified benchmarks) and Coding (3/3 ver The page publishes 14 verified benchmark rows for MiMo-V2.6-Pro, with an Agentic category weighted at 22% showing a score of 78.5 across 9 verified benchmarks. Coding is weighted at 20% with 3 verified benchmarks but score pending. The aggregator notes that no category currently has enough eligible evidence for a compa

LLM Stats ranks MiMo-V2.6-Pro at 18th overall on its composite LLM Stats Score (~50.0), with a blended price of $0.46 per million tokens. The model is benchmarked across multiple agent and coding evaluations, including a GDPval-AA Elo rating of 1673/3000 (rank 2), CyberGym score of 0.94 (rank 2), and Terminal-Bench 2.1 The aggregator contextualizes MiMo-V2.6-Pro's cost efficiency against competing models including Gemma 4 E4B ($0.024, score 13.8), GPT OSS 120B ($0.043, 28.7), DeepSeek-V4-Flash-0731 ($0.066, 43.8), DeepSeek-V4.1-Flash ($0.24, 51.4), and Claude Opus 5.5 ($4.76, 59.7), positioning MiMo-V2.6-Pro at $0.46 with a score of

Xiaomi officially released and open-sourced the MiMo-V2.6 series, comprising two natively multimodal models, Pro and Flash. The series was built around scaling reinforcement learning compute for recursive self-improvement on verifiable complex tasks, with MiMo-V2.6-Pro scoring 46 points on the Artificial Analysis Intel During the RL training phase, Xiaomi scaled compute across three axes: larger batches and higher throughput (up to 1,568 samples per update at 1M context length), more tasks and richer environments (coding, general agents, visual, cyber), and more grader compute. The excerpt notes MiMo-V2.6-Pro was likely one of the do