MiMo-V2.6-Flash

The eesel explainer confirms MiMo V2.6 is a family of open-weight, natively omnimodal models from Xiaomi's MiMo team released on September 22, 2026, with every checkpoint shipping a 1 million-token context window. Flash is explicitly named as a distinct checkpoint in the family, sharing the omnimodal input profile (tex The page frames the release under the tagline "Scaling Reinforcement Learning Toward Self-Improvement," with Xiaomi positioning the series as a step on a recursive self-improvement path via scaling RL compute on verifiable, complex tasks. The author notes that the technical report has enough concrete method behind it t

Xiaomi released and open-sourced the MiMo-V2.6 family on September 21, 2026, and MiMo-V2.6-Flash ships as one of the named variants alongside Pro and a 9B distillation. Per the page, Flash is 309B total parameters with 15B active, omnimodal across text, image, video, and audio, runs a 1M-token context, and was produced On pricing, Flash is listed at $0.14 per million input tokens and $0.28 per million output tokens on Xiaomi's own endpoint, unchanged from V2.5, while a faster Pro UltraSpeed serving variant costs $4.35/$8.70. Xiaomi also shipped the technical report, more than 7,000 RL environments, and the training framework with the

Benchable's page for Xiaomi's MiMo-V2.6-Flash, released September 21, 2026, describes a Mixture-of-Experts foundation model with 309B parameters and a 1M-token context, pricing at $0.14 input and $0.28 output per 1M tokens with $0.0028 cache reads. It reports a 99% reliability rate and competitive response times in the 53rd percentile across benchmarks. Highlighted benchmark accuracies include 100% on Hallucinations baseline, 99.5% General Knowledge, 99.0% Email Classification (90th percentile), 92.0% Reasoning, 91.0% Coding, 92.0% Mathematics, and 97.0% Ethics (29th percentile), with 69.0% Instruction Following. The model supports tools, structured outputs, reasoning, response format controls, and accepts text, image, audio, and video input while emitting text.

BenchLM's October 8, 2026 snapshot places MiMo-V2.6-Flash at capability 65.9/100, ranking 17th for Coding and 37th for Agentic across 122 to 146 models. Pricing is listed at $0.14 per 1M input and $0.28 per 1M output tokens with $0.003 cache reads and a 1M-token context window. Of 14 published benchmark rows, 9 are verified for Agentic and 3 for Coding; Reasoning, Multimodal, Knowledge, Multilingual, Instruction Following, and Math remain unmeasured. The page reports independent runtime speed as not measured, limiting speed-based conclusions while still marking the model as well-suited for software development and code generation.

MiMo-V2.6-Flash ranks 29th globally on RankLLMs with a composite score of 52.4, including a 4th-place OpenCode ranking and 67.2% on SWE-bench Verified. The model runs at 185 tokens per second with sub-220ms time-to-first-token and is reported as 4th in developer adoption on OpenCode with 9,486 daily active developers. Reported benchmark scores include 54.8% on GPQA Diamond, 62.0% on MATH-500, 32.5% on OSWorld, 52.0% on BrowseComp, 64.5% on Terminal-Bench 2.1, and a GDPVal-AA Arena Elo of 1560. The model is an open-weights, open-source release with a 1M-token context window, positioned as an ultra-fast multimodal option.

Xiaomi released MiMo-V2.6-Flash weights on Hugging Face on September 21, 2026, under an MIT license alongside MiMo-V2.6-Pro, with FP8 weights totaling 172.9 GB. The Flash checkpoint has 309B total and 15B active parameters across 48 layers (39 sliding-window and 9 global-attention) with a hidden size of 4096, and includes a 5-layer multi-token-prediction drafter that proposes up to 7 tokens per speculative decoding pass. Both MiMo-V2.6 models are natively multimodal via a 681M-parameter vision encoder and two audio encoders feeding the shared MoE backbone, supporting a 1M-token context. Xiaomi also published a 9B Qwen3.5-9B distillation, a technical report, RL code, and more than 7,000 training environments, and recommends SGLang with speculative decoding or vLLM with tensor parallelism of 4 for Flash.

MiMo-V2.6-Flash is an open-weights Xiaomi model scoring 38 on the Artificial Analysis Intelligence Index, placing it well above the 18 median among comparable models. It is priced at $0.14 per 1M input and $0.28 per 1M output tokens with a 98% cache discount, and it runs at 58 tokens per second while remaining reasonably priced relative to peers. The model uses a Mixture-of-Experts design with 309B total and 15B active parameters, ships under an MIT license, and supports a 1M-token context window with text and image input. A reasoning variant is noted, though the page evaluates the reasoning version; Artificial Analysis flags the model as slower than average and very verbose at 240M output tokens on the Intelligence Index.