MiMo-V2.6-Pro

Xiaomi released MiMo-V2.6-Pro and MiMo-V2.6-Flash on 22 September 2026 under the MIT license, with Pro scoring 46.32 on Artificial Analysis' Intelligence Index to top all open-weight models. The standout technical contribution is that Xiaomi also open-sourced the large-scale agentic reinforcement-learning infrastructure it spent roughly half a year building, including training environments and code, not just the model checkpoints. MiMo-V2.6-Pro handles text, images, video, and audio input within a 1-million-token context window, while Flash cuts API pricing to roughly a third of Pro without a major performance drop. Pro sits one point behind GPT-5.6 Sol (47) and five points behind Claude Opus 5 at max reasoning (51), while Alibaba's Qwen3.8 Max scores 45, Kimi K3 reaches 44, and DeepSeek V4.1 Flash lands at 39.

MiMo-V2.6-Pro is a Xiaomi-built foundation model released on September 22, 2026, scoring 46 on the Artificial Analysis Intelligence Index v4.3 and tying Grok 4.7, a new ceiling for open-weight models. The release was led by Luo Fuli, formerly of DeepSeek. Both Pro and Flash variants are available with open weights on Hugging Face under an MIT license, and the launch coincides with Alibaba's V900 chip unveiling, framing the moment as US-China frontier compression rather than a domestic-only alternative. MiMo-V2.6-Pro uses a frozen-router Mixture-of-Experts design with 1.02 trillion total and 42 billion active parameters, hybrid attention, a 5-layer MTP speculative decoder, and fully asynchronous GRPO training running 1,568 prompts across 16 rollouts per step. It maintains a 1-million-token context window with native multimodal capabilities. Vendor-reported benchmarks include DeepSWE v1.1 at 71.9%, Terminal Bench 2.1 at 89.9%, CyberGym at 94.0%, and AutomationBench v1.0.6 at 53.1%, outperforming GPT-5.6 Sol and Claude Fable 5 in several categories.

Xiaomi released MiMo-V2.6-Pro on 22 September 2026 as the top open-weights model on Artificial Analysis' Intelligence Index, scoring 46 points and tying with the newly released Grok 4.7. The flagship beats proprietary peers like Grok 4.6 (44) and Gemini 3.8 Flash (41), and surpasses open-weights rivals DeepSeek V4.1 Flash (39) and Pro (36). MiMo-V2.6-Pro is a 1.02-trillion-parameter mixture-of-experts model with 42 billion active parameters, a 1-million-token context window, and native text, image, audio, and video input. Xiaomi prices the API at $0.435 per million input tokens and $0.87 per million output tokens, with MIT-licensed weights on Hugging Face. It ships alongside the cheaper MiMo-V2.6-Flash, priced at $0.14/$0.28 per million tokens with the same context and multimodal support.

MiMo-V2.6-Pro is the highest-scoring open-weights model on Artificial Analysis' Intelligence Index at 46 points, tied with Grok 4.7, and roughly 45 times cheaper than Claude Opus 5 at max effort (51 points, $5.86 per task). Xiaomi's own benchmark table still trails Claude Opus 5 on 10 of 14 shared evaluations, with the widest gaps on ExploitBench, Terminal-Bench 4.0, and ProgramBench. The flagship uses a mixture-of-experts architecture with 1.02 trillion total and 42 billion active parameters, supports a 1,048,576-token context with 131,072-token max output, and accepts text, image, speech, and video input. Xiaomi lists the API model ID as mimo-v2.6-pro, prices it at $0.435 input and $0.87 output per million tokens, and publishes weights under MIT license on Hugging Face as MiMo-V2.6-Pro-RL.

OpenRouter's listing for xiaomi/mimo-v2.6-pro confirms MiMo-V2.6-Pro as the Xiaomi flagship foundation model with a 1M-token context window, native multimodal capabilities, and over 1T parameters, released September 21, 2026. Modalities include input and output, with pricing at $0.435 per million input tokens and $0.87 per million output tokens. The model is described as optimized for agentic workflows and long-horizon tasks, with top-tier performance across coding, visual, general, and research scenarios. The provider table lists two hosts: Xiaomi at $0.435 input, $0.87 output, $0.0036 cache read, 4.34s latency, 31 tokens per second throughput, and 99.89% uptime; and DeepInfra at the same pricing, 2.61s latency, 18 tokens per second throughput, and 91.35% uptime. Weighted-average prices across providers sit at $0.03207 per million input and $0.8696 per million output, reflecting heavy caching. Best observed latency is 2.61s P50 on DeepInfra and best throughput is 31 tokens per second P50 on Xiaomi.

MiMo-V2.6-Pro earns a composite Capability score of 74.2/100 on BenchLM, ranking 9th out of 210 tracked models as of September 29, 2026. It is particularly well-suited for software development, with a Coding rank of 9 of 143 at the 94th percentile based on 3 verified benchmarks. Agentic is the second-strongest category at rank 16 of 117, 87th percentile across 9 benchmarks, while Reasoning, Multimodal, Knowledge, Multilingual, Instruction Following, and Math are listed as Not measured. List pricing is reported at $0.43 per million input tokens and $0.87 per million output tokens, with a cached input price of $0.004 and a blended rate near $0.65. Speed sits at roughly 41 tokens per second against a field median of 92, with first-token latency around 53.23 seconds. Context window is 1 million tokens, and 14 verified benchmark rows are published on the model's tracker card for direct comparison against the wider catalog.