MiMo-V2.6-Flash-RL: Xiaomi's Efficient 309B Omnimodal Model
MiMo-V2.6-Flash-RL is Xiaomi's leaner 309B/15B-active MoE model, trading some benchmark points for speed against its Pro sibling.
What is MiMo-V2.6-Flash-RL?
MiMo-V2.6-Flash-RL is the efficiency-focused release in Xiaomi’s MiMo-V2.6 model family, a sparse Mixture-of-Experts (MoE) model with 309 billion total parameters but only 15 billion active per token. It’s the smaller sibling to MiMo-V2.6-Pro-RL, built to handle text, image, video, and audio in a single model with a 1 million token context window, while staying close to Pro-RL’s scores on coding, agentic, and cybersecurity benchmarks at a fraction of the inference cost.
TL;DR
- MiMo-V2.6-Flash-RL activates only 15B of its 309B total parameters per token, using a sparse MoE design with 256 routed experts and 8 activated per pass.
- The model is natively omnimodal, processing text, image, video, and audio through dedicated encoders (a 681M-parameter vision transformer and a combined 435M-parameter audio stack) rather than bolting vision or audio on afterward.
- It supports a 1 million token context window, aimed at long codebases, extended tool-use traces, and multi-session agent runs rather than single-shot chat.
- On most benchmarks Flash-RL lands within a few points of Pro-RL, but the gap widens sharply on harder cybersecurity tasks like ExploitBench (25.3 vs 47.9) and Terminal Bench 4.0 (28.8 vs 34.9).
- The model was trained with a single mixed reinforcement learning run across coding, general agent, visual, and cybersecurity tasks instead of separate RL passes per domain.
- Xiaomi pairs the RL run with a groupwise agentic grading system that ranks passing rollouts against each other, not just pass/fail, to keep improving quality after a task is already solvable.
- It ships with speculative decoding support (a 5-layer MTP drafter) and is deployable through SGLang, vLLM, or Xiaomi’s own API platform and desktop app.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How does MiMo-V2.6-Flash-RL differ from the Pro version?
Both models share the same MiMo-V2.6 architecture and training pipeline. The difference is scale and, by extension, cost. Pro-RL is the larger, higher-capacity checkpoint; Flash-RL is tuned to run faster and cheaper while giving up a controlled amount of accuracy.
Looking at the benchmark table Xiaomi published, the two models track closely on most agentic and coding tasks. On DeepSWE v1.1, Pro scores 71.9 and Flash scores 67.9. On MiMo Code Bench, it’s 63.2 versus 61.2. On the general agent side, AutomationBench actually favors Flash-RL by a hair (52.3 vs 53.1 is close enough to call a wash), and Toolathlon-Verified sits at 76.9 for Pro versus 73.6 for Flash.
The story changes in cybersecurity. Flash-RL actually edges out Pro on CyberGym (95.1 vs 94.0), but falls well behind on harder exploit-development tasks: ExploitBench drops from 47.9 (Pro) to 25.3 (Flash), and SEC Bench Pro drops from 66.3 to 47.5. That pattern suggests the smaller model’s activated parameter budget matters most on tasks that require deep, multi-step reasoning under adversarial conditions, where raw capacity has fewer places to hide.
What’s inside the architecture?
Flash-RL uses a hybrid attention backbone: 48 total transformer layers, with 39 using sliding window attention (SWA) and 9 using full global attention (GA). The first block runs global attention with a dense feed-forward layer; every block after that mixes local and global attention with sparse MoE feed-forward layers, no shared experts. Each token routes through 8 of 256 available experts.
The vision side runs on a dedicated 681-million-parameter vision transformer (MiMo ViT), 28 layers deep, split 24 SWA / 4 full attention, with a 2x2 spatial merge and 16x16 pixel patches. Audio runs through two components: a 308M-parameter AudioTokenizer using 20 residual vector quantization (RVQ) codebooks, plus a 127M-parameter patch encoder that compresses audio frames from 25 Hz down to 6.25 Hz before they hit the language backbone.
For inference speed, Flash-RL includes a 5-layer multi-token prediction (MTP) speculative decoder, described as DFlash-style, which predicts up to 7 subsequent tokens per forward pass for parallel verification. This is what lets a model with 309B total parameters run at speeds closer to a much smaller dense model, since only the 15B activated experts plus the lightweight drafter need to execute per step.
How was it trained?
The headline claim in Xiaomi’s technical report is “You Only RL Once”: rather than running separate reinforcement learning passes for coding, general agent tasks, visual tasks, and cybersecurity, the MiMo-V2.6 series trains all of them in one mixed RL run, with different tasks and harnesses mixed into the same batch. The idea is that skills learned in one domain transfer to harnesses the model never saw during training, and that mixing domains produces a more general agent than training each skill in isolation.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
The RL method is Group Relative Policy Optimization (GRPO), run fully asynchronously at large scale: 1,568 prompts times 16 rollouts per step, with billions of tokens processed per update. Binary pass/fail rewards can’t distinguish between a mediocre passing solution and an excellent one, so Xiaomi built two mechanisms to fix that. Groupwise Reward Synthesis (GRS) builds task-specific rubrics offline by contrasting rollouts within a group and fuses rubric quality with test outcomes. Groupwise Advantage Redistribution (GAR) does the ranking online, moving reward advantage toward the highest-quality trajectories rather than just any trajectory that passed. Xiaomi frames this as a self-improvement loop, since the model is graded against its own sampled outputs, and says it pushes the model toward shorter solution paths and fewer tokens per task.
After the mixed RL stage, the model goes through what Xiaomi calls Multi-Prefix Multi-Teacher On-Policy Distillation (MOPD2), which combines the model’s own autonomous rollouts with prefix-conditioned single-turn rollouts drawn from teacher trajectories and supervised fine-tuning demonstrations. The stated goal is extending capability into tasks that are hard to verify automatically, without needing to regenerate entire conversation histories for every training example.
Is MiMo-V2.6-Flash-RL worth deploying over Pro-RL?
For teams running high-volume coding agents, terminal automation, or visual coding tasks where the benchmark gap to Pro-RL is small (often 1-4 points), Flash-RL is a reasonable default: same architecture family, same training pipeline, lower activated parameter count, and presumably lower serving cost per token given 15B activated parameters versus whatever Pro-RL activates (Xiaomi’s public materials don’t disclose Pro-RL’s active parameter count).
For cybersecurity workloads specifically, especially exploit generation and hardened environments (ExploitBench, SEC Bench Pro), the gap to Pro-RL is large enough that Flash-RL probably isn’t the right choice if accuracy on those specific tasks matters more than throughput. It’s also worth noting that on several of these harder benchmarks, both MiMo models trail Claude Opus 5 and GPT-5.6 Sol, so neither MiMo checkpoint is a clear leader in adversarial security tasks against the strongest closed models.
How do you run it?
Flash-RL is released under the MIT license on Hugging Face and ModelScope, with weights available in both standard and fp8 formats. Xiaomi recommends serving it through SGLang using its published cookbook, with tensor parallelism, data parallelism, and EAGLE-based speculative decoding enabled for best throughput. A vLLM path is also documented, along with a pre-built Docker image tagged for the MiMo-V2.5 release line. Recommended sampling settings are temperature 1.0 and top_p 0.95. Beyond self-hosting, the model is accessible through Xiaomi’s AI Studio, MiMo Code, MiMo Desktop app, its own API platform, and OpenRouter.
Frequently Asked Questions
How many parameters does MiMo-V2.6-Flash-RL have?
It has 309 billion total parameters in a sparse Mixture-of-Experts architecture, with only 15 billion active per token, routing through 8 of 256 available experts.
What modalities does MiMo-V2.6-Flash-RL support?
It natively handles text, image, video, and audio in one model, using a dedicated 681M-parameter vision transformer and a combined 435M-parameter audio tokenizer and patch encoder.
How does Flash-RL compare to Pro-RL on benchmarks?
Flash-RL stays within a few points of Pro-RL on most coding and general-agent benchmarks, but falls noticeably behind on harder cybersecurity tasks like ExploitBench and SEC Bench Pro.
What context length does it support?
Up to 1 million tokens, intended for long codebases, extended agent tool-use traces, and multi-session runs rather than short single-turn chats.
How is MiMo-V2.6-Flash-RL licensed and where can it run?
It’s released under the MIT license on Hugging Face and ModelScope, deployable via SGLang or vLLM, and also available through Xiaomi’s AI Studio, MiMo Desktop, its API platform, and OpenRouter.
