MiMo-V2.6-Flash

Xiaomi released and open-sourced the MiMo-V2.6 series on September 22, 2026, including the Flash variant alongside the Pro flagship and a Pro-UltraSpeed tier. The family uses a sparse mixture-of-experts architecture with 1.02 trillion total parameters, 42 billion activated per token, and a 1-million-token context window, with native support for text, image, video, and audio inputs. These are vendor-reported figures. The MiMo-V2.6-Pro lead model reached an Artificial Analysis Intelligence Index score of 46, placing it first among open-source models. Xiaomi also released model weights, a technical report, training environments, and reinforcement-learning code alongside the launch. The Flash variant is positioned as a lower-cost, high-frequency counterpart within this open-weights lineup.

The MiMo-V2.6 series launched September 22, 2026, with both Pro and Flash variants open-sourced on Hugging Face under an MIT license. The Pro checkpoint uses a frozen-router MoE architecture with 1.02 trillion total parameters and 42 billion active, a 5-layer MTP speculative decoder, and fully asynchronous GRPO training across 1,568 prompts and 16 rollouts per step. These details are attributed to the Pro variant specifically. Vendor-reported benchmarks for MiMo-V2.6-Pro include 71.9% on DeepSWE v1.1, 89.9% on Terminal Bench 2.1, 94.0% on CyberGym, and 53.1% on AutomationBench v1.0.6, narrowing the gap with frontier proprietary systems. The team is led by Luo Fuli, formerly of DeepSeek. Flash availability is confirmed, though specific benchmark numbers for Flash are not detailed in this report.

Xiaomi has released MiMo-V2.6-Flash alongside the flagship MiMo-V2.6-Pro as part of the MiMo-V2.6 series, a pair of native multimodal language models with a 1-million-token context window supporting text, image, audio, and video input. According to VentureBeat, MiMo-V2.6-Flash is priced at $0.14 per million uncached in The model is positioned as a smaller, substantially cheaper sibling to MiMo-V2.6-Pro, targeting high-volume production workloads while retaining the same 1-million-token context window and native multimodal capabilities. Both variants are open-weight under an MIT license and available on Hugging Face, allowing develope

llm-stats.com lists MiMo-V2.6-Flash with a blended price of approximately $0.15 per million tokens and an LLM Stats Score of 45.4, placing it between DeepSeek-V4-Flash-0731 ($0.066, score 44.1) and DeepSeek-V4.1-Flash ($0.24, score 51.6) on its cost-efficiency chart. The page tracks Flash-specific benchmark results sou Performance-by-conversation-depth metrics on the page show MiMo-V2.6-Flash maintaining scores of 13.0, 12.7, and 12.4 across turns 1, 2-10, and 11-30 respectively (no data for turns 31+), suggesting relatively stable quality across longer conversations. The Quality Tracker shows a current sentiment at -0.6σ relative to

Xiaomi's first-party MiMo-V2.6 release page provides explicit vendor-reported benchmark scores for the Flash variant across multiple evaluations. On DeepSWE v1.1, Flash scored 67.9; on ProgramBench, 26.0; on MiMo Code Bench, 61.2; and on CyberGym, 95.1, leading the CyberGym leaderboard among reported models. Flash also scored 52.3 on Automation Bench v1.0.6, 73.6 on Toolathlon-verified, and 27.6 on Agents' Last Exam. The page further reports Flash scores of 28.8 on Terminal Bench 4.0, 61.2 on JobBench, and 71.5 on Visual Coding, positioning Flash as a balanced omnimodal model. The release framework scales reinforcement-learning compute on verifiable complex tasks to expand the capability frontier, with all three variants open-sourced as part of the September 22, 2026 launch. Benchmark figures are vendor-reported.

Xiaomi's official MiMo update log confirms three models in the V2.6 series released on September 22, 2026: mimo-v2.6-pro, mimo-v2.6-flash, and mimo-v2.6-pro-ultraspeed. The Flash variant is described as a full-modality, high-intelligence, low-cost reasoning model positioned as the best balance for high-frequency calls and large-scale tasks in professional workflows. The Pro-UltraSpeed variant targets up to 20x faster generation for latency-sensitive workloads. The same log shows the prior V2.5 generation from April 2026, including mimo-v2.5 and mimo-v2.5-pro with 1-trillion parameters, 42 billion activations, and 1-million-token context, establishing the architectural lineage carried into V2.6. A V2.5-asr speech recognition model was released June 2, 2026. Flash V2.6 extends the full-modality agent capabilities demonstrated by V2.5-pro into a cost-optimized tier.