MiMo-V2.6-Flash

Xiaomi's MiMo-V2.6 release includes MiMo-V2.6-Flash as one of two natively omnimodal models in the series, released with model weights, a technical report, training environments, and reinforcement-learning code. The series supports a 1-million-token context window, sparse mixture-of-experts architecture, and text, image, video, and audio inputs. On Artificial Analysis' Intelligence Index v4.3.2, the flagship MiMo-V2.6-Pro scores 46, placing it first among open-source models ahead of Z AI's GLM-5.3 (max) at 45 and Kimi K3 (max) at 44. Xiaomi cites MiMo-V2.6-Pro as the strongest open-source model to date, surpassing Kimi K3 and Qwen 3.8 Max, with the Pro checkpoint using 1.02 trillion total parameters and 42 billion activated per token.

Xiaomi's MiMo-V2.6 series, released September 22, 2026, includes MiMo-V2.6-Flash alongside the Pro flagship, with both variants now available as open weights on Hugging Face under an MIT license. The Pro variant uses a frozen-router Mixture-of-Experts architecture with 1.02 trillion total parameters and 42 billion active parameters, incorporating hybrid attention and a 5-layer MTP speculative decoder. The Pro model supports a 1-million-token context window with native multimodal capabilities and is trained with fully asynchronous GRPO across 1,568 prompts and 16 rollouts per step. Vendor-reported Pro results include 71.9% on DeepSWE v1.1, 94.0% on CyberGym (leading reported models), and 53.1% on AutomationBench v1.0.6, outpacing several US frontier counterparts on those benchmarks.

Xiaomi released MiMo-V2.6-Flash alongside its MiMo-V2.6-Pro flagship on September 21, 2026, positioning the smaller variant for high-volume production workloads. Per VentureBeat, Flash retains the same 1-million-token context window and native multimodal capabilities (text, image, audio, video) as Pro, while pricing it aggressively at $0.14 per million uncached input tokens and $0.28 per million output tokens via Xiaomi's API. The same report notes Flash is the second-cheapest major frontier model available via API by VentureBeat's assessment, aimed at indie developers and enterprises that want open-weight flexibility. MiMo-V2.6-Pro is MIT-licensed and downloadable from Hugging Face, and Xiaomi is named as the creator; Flash shares the family release though the report's headline anchors on Pro. Output speed for the family is measured at roughly 134 tokens per second with $0.13 per Artificial Analysis Intelligence Index task.

LLM Stats ranks MiMo-V2.6-Flash 31st on its composite LLM Stats Score, placing it in the "Average" tier across most categories. Capability standings include Coding at 22 of 273, Tool Calling at 29 of 200, Reasoning at 45 of 369, and Vision at 47 of 213, with Chat weaker at 115 of 121. On cost efficiency the model posts a blended price near $0.15 per million tokens and an LLM Stats Score around 45.7, positioning it between DeepSeek-V4-Flash and DeepSeek-V4.1-Flash on the chart. The page lists benchmark scores sourced from mimo.xiaomi.com, including CyberGym (rank 1, 0.95/1), Terminal-Bench 2.1 (rank 9, 0.88/1), and OSWorld-Verified, reinforcing strong agentic and computer-use evidence. A Quality Tracker panel shows performance variation around baseline across late September 2026, though the tracker has only 1 vote so far, so the signal is preliminary.

BenchLM's tracking page for MiMo-V2.6-Flash (data as of September 29, 2026) lists a capability score of 63.6/100, ranking the model 36th of 210 tracked models. Pricing matches the VentureBeat figures at $0.14 input / $0.28 output per million tokens, with $0.003 cached input and $0.21 blended, and confirms the 1-million-token context window. Independent runtime speed is not measured on this page. The page highlights Flash's strongest published evidence in the Agentic category (rank 41 of 117, 66th percentile) and Coding category (rank 44 of 143, 70th percentile), particularly useful for coding agents, browser research, and computer-use workflows. Reasoning, Multimodal, Knowledge, Multilingual, Instruction-Following, and Math categories show zero measured benchmarks, and BenchLM notes 14 published benchmark rows across the tracked categories.

Xiaomi released and open-sourced the MiMo-V2.6 series on September 22, 2026, with MiMo-V2.6-Flash positioned as a natively omnimodal variant balancing intelligence, efficiency, and cost alongside the flagship Pro. The series is part of Xiaomi's exploration of scaling reinforcement learning on verifiable, complex tasks to expand the capability frontier through exploration and feedback. MiMo-V2.6-Flash shares the series' 1-million-token context window and native multimodal inputs (text, image, video, audio). Xiaomi-reported benchmarks include 67.9 on DeepSWE v1.1, 26.0 on ProgramBench, 61.2 on the MiMo Code Bench, and 73.6 on Toolathlon-verified, trailing the Pro variant on coding and agent evaluations. A Pro-UltraSpeed variant delivers up to 20x faster output at the same quality for users needing extreme generation speed.