MiMo-V2.6-Pro

Forkast reports that MiMo-V2.6-Pro uses a frozen-router Mixture-of-Experts architecture with 1.02 trillion total parameters and 42 billion active parameters, incorporating hybrid attention mechanisms and a 5-layer MTP speculative decoder. Training used fully asynchronous GRPO with 1,568 prompts across 16 rollouts per s On vendor-reported benchmarks, MiMo-V2.6-Pro achieved 71.9% on DeepSWE v1.1, 89.9% on Terminal Bench 2.1 (surpassing Claude Opus 5 and GPT-5.6 Sol), 94.0% on CyberGym (leading all reported models), and 53.1% on AutomationBench v1.0.6. The article also contextualizes the release alongside the Alibaba V900 chip unveiling

Xiaomi officially released and open-sourced the MiMo-V2.6 series, comprising two native fully multimodal models (Pro and Flash), on September 22, 2026. The release documents a 6-day live RL training phase as part of Xiaomi's recursive self-improvement exploration, building on verifiable complex tasks and scaling up RL The V2.6 series adopts the same API pricing as V2.5, with Xiaomi framing MiMo-V2.6-Pro as setting a new cost-performance record among domestic large language models at 1/20 to 1/60 the price of overseas models at comparable intelligence. MiMo-V2.6-Pro is described as one of the domestic open-source models with the larg

Xiaomi announced and open-sourced the MiMo-V2.6 series on September 22, 2026, positioning it as a step along the RSI path of scaling RL compute on verifiable, complex tasks. The lineup has two natively omnimodal models: MiMo-V2.6-Pro as the most capable to date, and MiMo-V2.6-Flash tuned for the balance of intelligence, efficiency, and cost, plus MiMo-V2.6-Pro-UltraSpeed. The UltraSpeed variant is claimed to deliver up to 20x faster output speed at the same quality for extreme generation-speed needs. Xiaomi's own tables list MiMo-V2.6-Pro scoring 71.9 on DeepSWE v1.1, 1673 on GDPVal 2.1, 72.3 on MiMo Visual Coding, and 94.0 on CyberGym, placing it near top models from OpenAI, Anthropic and DeepSeek on these agentic and coding suites.

Xiaomi's technical report (arXiv:2610.11959v1, submitted 8 Oct 2026) introduces the MiMo-V2.6 series as an omni-modal family built around scaling reinforcement-learning compute, with MiMo-V2.6-Pro tracked separately from MiMo-V2.6-Flash across RL training. The report frames recursive self-improvement as depending on agents that couple models with interactive environments to produce multi-step trajectories and feedback. Training details include mid-training on a broad multimodal corpus before RL, and an asynchronous setup consuming 1,568 samples and 2.7–3.7B tokens per step at context lengths up to 1M. The team freezes the MoE router and adds a multi-layer defense against reward hacking, while releasing training dynamics, RL environments, and the RL framework for reproduction.

The MiMo-V2.6 technical report describes an omni-modal family that scales reinforcement learning compute, beginning with mid-training on a broad multimodal corpus on a pretrained hybrid-SWA architecture to widen the exploration space. RL is scaled along three dimensions: larger batches and higher throughput, more diverse code, general, visual, and cyber environments, and additional grader compute via groupwise agentic grading for long-horizon tasks. Stability at scale is maintained by freezing the MoE router and adding a multi-layer defense against reward hacking. The infrastructure includes a unified trajectory representation, high-concurrency multi-framework rollout, decoupled control and data planes, and training inference consistency, with asynchronous runs consuming 1,568 samples and roughly 2.7 to 3.7 billion tokens per step at context lengths up to 1M. Training dynamics, RL environments, and the RL framework are open-sourced.

Xiaomi officially released and open-sourced the MiMo-V2.6 series, including MiMo-V2.6-Pro and MiMo-V2.6-Flash as native fully multimodal models. The release emphasizes scaled reinforcement learning for recursive self-improvement, including six days of live RL training backed by roughly six months of foundational research and engineering work. MiMo-V2.6-Pro reportedly scores 46 on the Artificial Analysis Composite Intelligence Index, which Xiaomi says exceeds Kimi K3 and Qwen3.8 Max, though it still trails leading closed models. The V2.6 series retains V2.5-series API pricing, and Xiaomi says the retired V2.5-Pro and V2.5 model IDs stopped working on October 21, 2026, requiring migration to V2.6.