MiMo-V2.6 Flash-RL vs Pro-RL: Which Xiaomi Model to Run
Xiaomi's MiMo-V2.6 Flash-RL (309B/15B active) vs Pro-RL (1T/42B active): architecture, benchmarks, and deployment tradeoffs compared.
What’s the difference between MiMo-V2.6 Flash-RL and Pro-RL?
MiMo-V2.6-Pro-RL is Xiaomi’s flagship open-weight model at 1.02 trillion total parameters with 42 billion active per token. Flash-RL is the smaller sibling built for lighter deployment, at 309 billion total parameters with 15 billion active. Both share the same training pipeline, the same 1M-token context window, and the same omnimodal input support (text, image, video, audio). The gap between them shows up in layer count, hidden size, and expert pool size, which translates into a real but modest benchmark gap and a much bigger difference in what hardware you need to run them.
TL;DR
- Pro-RL activates roughly 2.8x more parameters per token than Flash-RL (42B vs 15B), while total parameters differ by about 3.3x (1.02T vs 309B).
- Both models share the same architecture family: hybrid sliding-window/global attention, sparse MoE without shared experts, a 5-layer MTP speculative decoder, and identical vision (681M) and audio (308M + 127M) encoders.
- Benchmark scores favor Pro-RL almost everywhere, but the margins are usually single digits, not the kind of gap you’d expect from a 3x parameter difference.
- The gap widens sharply on cybersecurity benchmarks, where Flash-RL falls well behind Pro-RL on ExploitBench and SEC Bench Pro, suggesting harder adversarial tasks benefit more from scale.
- Flash-RL actually edges out Pro-RL on CyberGym (95.1 vs 94.0), a reminder that bigger doesn’t always win on every eval.
- Deployment requirements scale roughly with parameter count: Pro-RL’s reference SGLang config uses 16-way tensor parallelism across 2 nodes, while Flash-RL runs on 8-way tensor parallelism on a single node.
- Both use identical RL training methodology (GRPO, Groupwise Agentic Grading, MOPD2 distillation), so the difference is capacity, not technique.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How do the two architectures actually compare?
The two models are built from the same blueprint but scaled down uniformly. Pro-RL’s backbone has 70 transformer layers (60 sliding-window attention, 10 global attention), a hidden size of 6144, and a pool of 384 routed experts with 8 activated per token. Flash-RL trims this to 48 layers (39 SWA, 9 global), a hidden size of 4096, and 256 routed experts, still activating 8 per token.
Both models keep the same head dimensions (192 for QK, 128 for V) and the same sliding window size of 128, so the attention mechanics are identical, just applied across fewer layers and a smaller hidden dimension in Flash-RL. Neither model uses shared experts in its MoE FFNs, and both open with a dense global-attention block before switching into the interleaved SWA/GA pattern.
The multimodal stack is untouched between the two. Vision encoding runs through the same 681M-parameter MiMo ViT (28 layers, 24 SWA + 4 full), and audio runs through the same 308M AudioTokenizer paired with a 127M patch encoder. If you’re building a pipeline that needs image or audio input, switching from Pro-RL to Flash-RL costs you nothing on the multimodal side. The savings come entirely from the language backbone.
Both models also ship the same 5-layer MTP (multi-token prediction) speculative decoder, predicting 7 tokens ahead per forward pass for parallel verification. This is a shared efficiency feature, not something exclusive to the smaller model.
Is the benchmark gap between them large?
Not as large as the parameter difference would suggest. Across code agent benchmarks, Pro-RL leads Flash-RL by small margins: 71.9 vs 67.9 on DeepSWE v1.1, 63.2 vs 61.2 on MiMo Code Bench, and a near-tie on ProgramBench (26.5 vs 26.0). On general agent tasks the pattern holds: Pro-RL scores 76.9 on Toolathlon-Verified against Flash-RL’s 73.6, and 82.0 vs 80.8 on OSWorld-Verified. On Terminal Bench 2.1 the two are within 2.3 points of each other (89.9 vs 87.6).
Cybersecurity is where the story changes. On ExploitBench, Pro-RL scores 47.9 against Flash-RL’s 25.3, nearly double. SEC Bench Pro shows a similar split: 66.3 vs 47.5. These are the benchmarks with the most adversarial, open-ended tasks in the suite, and they appear to reward the extra activated parameters more than agentic or coding tasks do. Interestingly, Flash-RL actually beats Pro-RL on CyberGym (95.1 vs 94.0) and comes close on MiMo Cyber Bench (77.2 vs 80.2), so the cybersecurity story isn’t uniformly in Pro-RL’s favor, it’s specifically the harder exploit-generation benchmarks where scale matters.
For context, both MiMo-V2.6 checkpoints post large improvements over the prior generation, MiMo-V2.5-Pro, which scored dramatically lower across nearly every benchmark listed (19.0 on DeepSWE v1.1, 0.0 on MiMo Cyber Bench, 1.5 on Terminal Bench 4.0). Whatever changed in the V2.6 training pipeline, both the Pro and Flash checkpoints benefited from it.
Against external models, Pro-RL trades wins and losses with Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5 depending on the benchmark, sometimes leading (AutomationBench v1.0.6, MiMo VisualCoding is close), sometimes trailing (Terminal Bench 4.0, ExploitBench). Flash-RL is more consistently behind these larger external models but not by wide margins on most agentic tasks.
What does deployment actually cost for each model?
This is where the parameter counts stop being an abstraction. Xiaomi’s reference SGLang command for Pro-RL specifies --tp 16 --dp 2 --nnodes 2, meaning the recommended setup spans two nodes with 16-way tensor parallelism and 2-way data parallelism, plus expert parallelism (--ep 16) with a DeepEP backend for MoE all-to-all communication. That’s a multi-node cluster.
Flash-RL’s reference command uses --tp 8 --dp 2 on a single node, no --ep flag, no multi-node addressing. The vLLM recipes follow the same pattern: Pro-RL recommends --tensor-parallel-size 8, Flash-RL drops to --tensor-parallel-size 4.
Both models use the same speculative decoding setup (EAGLE, 3 steps, topk 1, 4 draft tokens, multi-layer EAGLE enabled) and the same reasoning and tool-call parsers (mimo). The serving stack and inference tricks are identical, meaning the deployment complexity difference comes down almost entirely to how many GPUs and nodes you need to fit the weights and KV cache, not to different software requirements.
Which one should you actually run?
If you’re running agentic coding or general automation workloads and have access to a multi-node GPU cluster, Pro-RL’s benchmark lead is real, if modest, across code and general agent tasks. If your workload touches cybersecurity research, exploit generation, or adversarial red-teaming tasks, the gap becomes large enough that it’s worth the extra compute.
If you’re deploying on a single node, want to serve at lower latency, or your use case leans toward general agentic tasks, coding assistance, or visual coding rather than adversarial security work, Flash-RL gets you most of the capability at roughly a third of the activated parameters and about half the deployment footprint. The MiMo VisualCoding scores (72.3 vs 71.5) and general agent scores are close enough that Flash-RL is a reasonable default unless you have a specific reason to need Pro-RL’s extra headroom.
Both models are available through the same channels: Xiaomi’s AI Studio, MiMo Code, MiMo Desktop, the Xiaomi MiMo Open Platform API, and OpenRouter, alongside direct weight downloads from Hugging Face and ModelScope.
Frequently Asked Questions
How many parameters does MiMo-V2.6-Flash-RL activate per token?
Flash-RL activates 15 billion parameters per token out of 309 billion total, using a sparse Mixture-of-Experts architecture with 8 of 256 routed experts active at once.
How many parameters does MiMo-V2.6-Pro-RL activate per token?
Pro-RL activates 42 billion parameters per token out of 1.02 trillion total, drawing from a pool of 384 routed experts with 8 active per forward pass.
Do Flash-RL and Pro-RL support the same context length?
Yes, both models support up to 1 million tokens of context, and both use the same hybrid sliding-window and global attention design plus the same 5-layer MTP speculative decoder.
Which model performs better on cybersecurity tasks?
Pro-RL leads clearly on ExploitBench and SEC Bench Pro, but Flash-RL actually scores higher on CyberGym (95.1 vs 94.0), so the advantage isn’t uniform across every cybersecurity benchmark.
Can I run Flash-RL on a single GPU node?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
Xiaomi’s reference SGLang deployment for Flash-RL uses 8-way tensor parallelism on a single node, versus Pro-RL’s 16-way tensor parallelism across two nodes, making Flash-RL considerably lighter to deploy.
