MiMo-V2.6-Pro-RL: Xiaomi's 1T-Parameter Agentic Model, Explained
Xiaomi's MiMo-V2.6-Pro-RL is a 1T-parameter MoE model trained with large-scale RL. Here's how it compares to Claude Opus 5 and GPT-5.6.
What is MiMo-V2.6-Pro-RL?
MiMo-V2.6-Pro-RL is an open-weight, trillion-parameter mixture-of-experts (MoE) language model from Xiaomi’s MiMo team, built specifically for agentic work: coding, tool use, browser and terminal automation, and cybersecurity tasks. It has 1.02 trillion total parameters with only 42 billion active per token, a 1 million token context window, and native support for text, image, video, and audio. On several coding and general-agent benchmarks it lands close to or ahead of Claude Opus 5 and GPT-5.6, according to Xiaomi’s own published evaluation table.
TL;DR
- MiMo-V2.6-Pro-RL is a sparse MoE model with 1.02T total parameters and 42B activated per token, keeping inference cost closer to a 42B dense model despite the massive parameter count.
- The context window reaches 1M tokens, enough to hold entire repositories, long tool-call traces, or multi-session agent transcripts in a single pass.
- Xiaomi trained it with one large mixed reinforcement learning run across coding, general agents, vision, and cybersecurity tasks instead of separate RL runs per domain, aiming for skills that transfer across harnesses the model never saw during training.
- The reward system uses groupwise agentic grading rather than simple pass/fail scoring, comparing rollouts against each other to rank quality and push the model toward shorter, more efficient solution paths.
- On benchmarks like AutomationBench and Terminal Bench 2.1, MiMo-V2.6-Pro scores competitively with Claude Opus 5 and GPT-5.6, though it trails both models clearly on exploit-focused cybersecurity tests like ExploitBench.
- The model ships with native omnimodal encoders: a 681M-parameter vision transformer and a combined audio stack of roughly 435M parameters, plus a 5-layer speculative decoder for faster generation.
- It’s deployable today via SGLang or vLLM, and also available through Xiaomi’s own API platform, AI Studio, MiMo Desktop, and OpenRouter.
How is MiMo-V2.6-Pro-RL built?
The model uses a hybrid attention backbone: 70 transformer layers total, 60 using sliding window attention (SWA) and 10 using full global attention (GA), with 384 routed experts of which 8 activate per token. The first block runs global attention with a dense feed-forward network; every block after that mixes local and global attention layers, each backed by sparse MoE feed-forward networks with no shared experts.
Three encoders feed into this backbone to make it natively multimodal. The vision encoder, called MiMo ViT, has 681 million parameters across 28 layers (24 sliding-window, 4 full attention) and processes video in 2x16x16 patches. Audio comes through two components: a 308M-parameter tokenizer using 20 residual vector quantization codebooks, and a 127M-parameter patch encoder that compresses audio from 25 Hz down to 6.25 Hz. A 5-layer speculative decoder, described in the model card as DFlash-style, predicts seven tokens ahead per forward pass to speed up inference.
None of this is unusual in isolation. MoE architectures, sliding-window attention, and speculative decoding are all established techniques. What’s notable is the scale at which Xiaomi combined them into a single omnimodal checkpoint with a million-token context, and the training approach layered on top.
What makes the reinforcement learning here different?
Most RL-tuned models are trained separately for each skill: one run for coding, another for tool use, another for safety-relevant tasks. MiMo-V2.6-Pro-RL was trained with what Xiaomi calls “You Only RL Once”: a single mixed reinforcement learning run spanning coding, general agent tasks, visual tasks, and cybersecurity, with different task types and harnesses mixed into the same training batches. The stated goal is for skills learned in one domain to reinforce performance in another, and for strategies to generalize to agent harnesses the model never encountered during training.
The RL itself runs on Group Relative Policy Optimization (GRPO) at large scale, asynchronously, with batches of 1,568 prompts times 16 rollouts per step, amounting to billions of tokens per update according to the model card.
The more interesting piece is how rewards get computed. Binary pass/fail signals can’t distinguish a clean solution from a clunky one that happens to pass. Xiaomi’s answer is “groupwise agentic grading”: an automated grader compares rollouts against each other within the same group rather than scoring each in isolation. Two mechanisms do this work. Groupwise Reward Synthesis builds task-specific scoring rubrics offline by contrasting different rollouts, then blends rubric quality with actual test outcomes. Groupwise Advantage Redistribution ranks passing trajectories online and shifts more training signal toward the better ones. Because the grading judges the model’s own outputs against each other, Xiaomi describes this as a self-improvement loop, one that also pushes the model toward shorter, more token-efficient solutions over training.
To keep that loop from producing a model that games its own grader, Xiaomi added environment hardening, adversarial screening, and cross-checks from separate verifiers during training. After the mixed RL phase, a distillation step called MOPD2 (Multi-Prefix Multi-Teacher On-Policy Distillation) extends the resulting capabilities to tasks that are harder to verify automatically, by mixing the model’s own rollouts with rollouts conditioned on teacher and supervised fine-tuning demonstrations.
How does it perform against Claude Opus 5 and GPT-5.6?
Xiaomi’s published benchmark table puts MiMo-V2.6-Pro roughly in the same tier as Claude Opus 5 and GPT-5.6 on several agent and coding tests, while trailing on others.
On coding benchmarks, MiMo-V2.6-Pro scores 71.9 on DeepSWE v1.1 versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6, and 63.2 on MiMo’s own Code Bench versus 68.6 and 59.3 respectively. On ProgramBench it falls further behind at 26.5 against Claude Opus 5’s 37.0.
General agent tasks show a stronger showing. MiMo-V2.6-Pro leads on AutomationBench v1.0.6 with 53.1 against 50.3 (Claude Opus 5) and 45.8 (GPT-5.6). It’s close behind on Toolathlon-Verified (76.9 vs. 80.6 and 74.9) and essentially tied on Agents’ Last Exam (31.6, matching Claude Opus 5 exactly, ahead of GPT-5.6’s 30.8). On Terminal Bench 2.1 it edges out both at 89.9 versus 89.1 and 88.8. On Terminal Bench 4.0, a harder variant, it drops to 34.9 against Claude Opus 5’s 49.0.
Cybersecurity is where the gap widens most. MiMo-V2.6-Pro posts strong numbers on Xiaomi’s own CyberGym (94.0) and MiMo Cyber Bench (80.2), but on externally-flavored exploit tests it lags: 47.9 on ExploitBench versus 70.0 (Claude Opus 5) and 78.5 (GPT-5.6), and 17.8 on ExploitGym versus 22.1 and 30.3.
The pattern across the table: MiMo-V2.6-Pro is genuinely competitive on general agent and automation tasks, roughly even with the leading closed models on some coding benchmarks, and clearly behind on offensive-security-style exploit benchmarks.
Is MiMo-V2.6-Pro-RL worth running?
For teams that want an open-weight model with agentic performance in the same range as Claude Opus 5 and GPT-5.6, it’s a serious option, provided the hardware is available. Running the full model requires substantial multi-GPU infrastructure. Xiaomi’s own deployment example for SGLang specifies tensor parallelism of 16 across 2 nodes with expert parallelism and DeepEP MoE routing, and the vLLM example uses 8-way tensor parallelism. This is not a model for a single consumer GPU; it targets serious inference clusters.
A smaller sibling, MiMo-V2.6-Flash-RL, is also available for teams that need lower latency or smaller footprint at some cost to benchmark scores (Flash trails Pro by roughly 2 to 5 points on most agent benchmarks, and more on the harder cybersecurity tests).
Both models are released under an MIT license and available on Hugging Face and ModelScope, along with hosted access through Xiaomi’s API platform, AI Studio, MiMo Desktop app, and OpenRouter, for teams that don’t want to self-host.
Frequently Asked Questions
How many parameters does MiMo-V2.6-Pro-RL have?
It has 1.02 trillion total parameters in a mixture-of-experts architecture, but only 42 billion parameters activate per token, since just 8 of 384 routed experts fire for any given input.
Does MiMo-V2.6-Pro-RL beat Claude Opus 5 and GPT-5.6?
It’s mixed. Xiaomi’s own benchmarks show it ahead on some general-agent tests like AutomationBench and Terminal Bench 2.1, roughly tied on others like Agents’ Last Exam, and behind on coding tests like ProgramBench and cybersecurity exploit benchmarks like ExploitBench.
What context length does it support?
Up to 1 million tokens, which the model card says is meant for long repositories, extended tool-call traces, and multi-session agent runs.
Can I run MiMo-V2.6-Pro-RL on a single GPU?
No. Xiaomi’s deployment examples use 8 to 16-way tensor parallelism across multiple GPUs and nodes via SGLang or vLLM. A smaller variant, MiMo-V2.6-Flash-RL, is available for lighter deployment needs.
What license does MiMo-V2.6-Pro-RL use?
It’s released under the MIT license and available as open weights on Hugging Face and ModelScope, alongside hosted access through Xiaomi’s API and app platforms.
