MiMo-V2.6-Pro

Forkast reports Xiaomi's MiMo-V2.6-Pro scoring 46 on Artificial Analysis Intelligence Index v4.3 — tying Grok 4.7 and setting a new ceiling for open-weight models. The model uses a frozen-router Mixture-of-Experts architecture with 1.02T total / 42B active parameters, hybrid attention, a 5-layer MTP speculative decoder Vendor-reported benchmarks cited: DeepSWE v1.1 71.9% (trailing Claude Opus 5's 74.0% but ahead of Claude Fable 5), Terminal Bench 2.1 89.9% (surpassing Claude Opus 5 and GPT-5.6 Sol), CyberGym 94.0% (leading all reported models), and AutomationBench v1.0.6 53.1%. The article frames the release alongside Alibaba's annou

The BenchLM aggregator profile for MiMo-V2.6-Pro reports 14 published benchmark rows with verified coverage only in Agentic (9/9) and Coding (3/3) categories; Reasoning, Multimodal, Knowledge, Multilingual, Instruction Following, and Math remain unmeasured, and the model is therefore marked "Not ranked." Reported opera Because most capability categories lack sufficient published evidence, BenchLM does not assign comparative ranks for MiMo-V2.6-Pro and notes that prospective users should validate before choosing. The profile functions as an aggregation layer over third-party benchmark sources rather than a primary measurement, and its

Latent Space's AINews roundup covers Xiaomi's launch of MiMo-V2.6-Pro (1T total / A42B active) as the top open-weights model on the Artificial Analysis Intelligence Index (score 46) at $0.13 per task, trained for roughly $3M. The piece highlights that Xiaomi is positioning itself as a new Chinese frontier lab, with the On the technical training side, the article summarizes Xiaomi's RL scaling along three axes: larger batches and higher throughput (1,568 samples per update, up to 1M context length, 3.5–3.7B tokens per step on a fully asynchronous architecture); more tasks and richer environments spanning coding, general agents, visual

Xiaomi released the MiMo-V2.6 series on 22 September 2026, with MiMo-V2.6-Pro as the flagship open-weights model. According to Artificial Analysis, MiMo-V2.6-Pro scores 46 on the Intelligence Index, the highest ever for an openly available model, at roughly US$0.13 per task. The article details three variants — Pro (1. The piece frames MiMo-V2.6-Pro as a new Pareto-frontier entrant for agentic automation (tool use, terminal work, computer use, multi-step workflows). It cites Xiaomi's launch claim of parity with leading frontier models on agentic benchmarks at a fraction of closed-lab pricing, and is aimed at a business-decision-maker