# MiMo-V2.6-Flash-RL ![Organization](https://img.shields.io/badge/Org-Xiaomi-FF6900) ![License](https://img.shields.io/badge/License-MIT-2da44e) ![Params](https://img.shields.io/badge/Params-309B%20(15B%20active)-8957e5) ![Context](https://img.shields.io/badge/Context-1M%20tokens-blue) ![Released](https://img.shields.io/badge/Released-Sep%202026-555) > **Quick answer:** MiMo-V2.6-Flash-RL is the efficiency-balanced checkpoint of Xiaomi's MiMo-V2.6 series: a 309B-parameter sparse MoE (15B active) with text, image, video and audio input and a 1M-token context, released September 21, 2026 under the MIT license. It reports 87.6 on Terminal-Bench 2.1, 67.9 on DeepSWE v1.1 and 73.6 on Toolathlon-Verified — close to its 1.02T-parameter sibling at roughly a third of the active parameters. ## At a Glance **Where it leads:** Near-flagship agentic coding and computer use for its size: OSWorld-Verified 80.8, Terminal-Bench 2.1 87.6 and DeepSWE v1.1 67.9. **Where it lags:** Terminal-Bench 4.0 (28.8), ExploitBench (25.3) and SEC-Bench Pro (47.5) are well below the Pro checkpoint and the closed frontier models. **Best for:** Cost-efficient self-hosted agents and long-context coding workloads that still need omnimodal input. ## What MiMo-V2.6-Flash-RL Is It shares the MiMo-V2.6 recipe with [MiMo-V2.6-Pro-RL](/models/xiaomi/mimo-v2-6-pro-rl): one mixed asynchronous GRPO run across coding, general agents, vision and cybersecurity, groupwise agentic grading, and multi-teacher on-policy distillation. The 48-layer backbone interleaves 39 sliding-window layers with 9 global-attention layers, 256 routed experts (8 active), a 681M-parameter vision encoder, audio encoders and a 5-layer multi-token-prediction drafter. Weights are open (FP8, MIT license) and the model is available via the Xiaomi MiMo API platform, AI Studio, MiMo Code and OpenRouter. ## Specifications | Field | Value | |---|---| | Organization | Xiaomi | | Hugging Face | [XiaomiMiMo/MiMo-V2.6-Flash-RL](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | | Architecture | Sparse MoE, 309B total / 15B active | | Context length | 1M tokens | | Modalities | Text, image, video, audio | | License | MIT | | Release date | 2026-09-21 | ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [DeepSWE](/benchmarks/datacurve/deepswe) (v1.1) | 67.9 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | | [Terminal-Bench 4.0](/benchmarks/anthropic/terminal-bench-40) | 28.8 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | | [Terminal-Bench 2.1](/benchmarks/tbench/terminalbench-21) | 87.6 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | | [Toolathlon-Verified](/benchmarks/hkust-nlp/toolathlon-verified) | 73.6 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | | [OSWorld-Verified](/benchmarks/osworld/osworld-verified) | 80.8 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | | [JobBench](/benchmarks/li-et-al/jobbench) | 61.2 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | | [ExploitBench](/benchmarks/seunghyun-lee-david-brumley/exploitbench) | 25.3 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | | [SEC-Bench Pro](/benchmarks/sec-bench-pro) | 47.5 | [Model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Flash-RL) | 2026-09 | *Self-reported by Xiaomi; not independent Benchgen measurements. Also reported but not added: ProgramBench 26.0 and Agents' Last Exam 27.6 (metric differs), AutomationBench v1.0.6 52.3 (version-mismatch precedent), CyberGym 95.1 and ExploitGym 6.0 (scale differs), and Xiaomi-internal MiMo Code Bench, MiMo Cyber Bench and MiMo VisualCoding.* ## Frequently Asked Questions

What is MiMo-V2.6-Flash-RL?

Xiaomi's efficiency-balanced MiMo-V2.6 checkpoint: a 309B-parameter omnimodal MoE (15B active) with a 1M-token context, released September 21, 2026 under the MIT license.

How is it different from MiMo-V2.6-Pro-RL?

Pro-RL is the 1.02T-parameter (42B active) flagship; Flash-RL is about a third of the size with similar scores on most agentic benchmarks but a clear gap on cybersecurity tasks.

Is MiMo-V2.6-Flash-RL open weights?

Yes, MIT-licensed weights are on Hugging Face, and it is also served through the Xiaomi MiMo API platform and OpenRouter.

MiMo-V2.6-Flash-RL — Training jobs

Training jobs