MiMo-V2.6-Flash

Xiaomi released the MiMo-V2.6 series on September 22, 2026, shipping both Pro and Flash variants as open weights on Hugging Face under an MIT license. The release is led by Luo Fuli, formerly of DeepSeek, and positions MiMo-V2.6-Pro at a 46 score on the Artificial Analysis Intelligence Index v4.3, tying Grok 4.7 and establishing a new ceiling for open-weight models. The MiMo-V2.6 series 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. The design supports a 1-million-token context window with native multimodal capabilities, trained via fully asynchronous GRPO across 1,568 prompts and 16 rollouts per step. Both Pro and Flash variants are immediately accessible to global builders on Hugging Face.

VentureBeat reports MiMo-V2.6-Flash launched alongside the flagship Pro on September 21, 2026, with Xiaomi pricing it at $0.14 per million uncached input tokens and $0.28 per million output tokens. VentureBeat calls Flash the second-cheapest major frontier model available over API among those assessed, targeting high-volume production workloads. Per VentureBeat, Flash retains the same 1-million-token context window and native multimodal capabilities as Pro, accepting text, image, audio, and video input. Both variants ship open-weight under an MIT license on Hugging Face, with Xiaomi's own API as an alternative access path.

MiMo-V2.6-Flash from Xiaomi ranks 31st on the LLM Stats Score composite leaderboard, with capability tiers of A in Coding (22 of 273), B in Tool Calling (29 of 200) and Reasoning (45 of 369), and C in Chat (115 of 121). The model is priced at $0.15 blended per million tokens, positioned between DeepSeek-V4-Flash-0731 ($0.066) and DeepSeek-V4.1-Flash ($0.24) on the cost-efficiency chart, sourced from llm-stats.com benchmark data on Sep 28, 2026. Vendor-reported benchmark scores from mimo.xiaomi.com show MiMo-V2.6-Flash achieving 0.95 on CyberGym (rank 1), 0.88 on Terminal-Bench 2.1 (rank 9), and a visible OSWorld-Verified result on the llm-stats scorecard. The Quality Tracker currently shows zero community votes as of the Sep 22, 2026 baseline, indicating the model is too new for independent community quality signals. Performance-across-conversation-depth tracking is active with a baseline established Sep 22–27.

The BenchLM catalog page for MiMo-V2.6-Flash records pricing of $0.14 per million input tokens, $0.28 per million output tokens, and $0.003 per million cached input tokens, with a 1-million-token context window. It lists 14 published benchmark rows across categories such as Agentic, Coding, and Multimodal, though speed is not independently measured. BenchLM notes MiMo-V2.6-Flash has published evidence in agentic and coding categories but lacks sufficient eligible rows for a public rank in Reasoning, Multimodal, Knowledge, and several others. Data is timestamped September 23, 2026, and category scores are described as source-linked rather than independently verified.

Xiaomi released and open-sourced MiMo-V2.6-Flash on September 22, 2026, as part of the MiMo-V2.6 series alongside MiMo-V2.6-Pro. Xiaomi describes Flash as a natively omnimodal model that balances intelligence, efficiency, and cost, with a separate Pro-UltraSpeed variant delivering up to 20x faster output at the same quality. According to Xiaomi's official page, MiMo-V2.6-Flash shares the 1-million-token context window and native multimodal input support of the series. Xiaomi-reported benchmark rows include DeepSWE v1.1 at 67.9, MiMo Code Bench at 61.2, Toolathlon at 73.6, Agents' Last Exam at 27.6, Terminal Bench 4.0 at 28.8, Visual Coding at 71.5, and CyberGym at 95.1.