MiMo V2.6: Xiaomi's Open Model Trained Live for $3.5M
Xiaomi's MiMo V2.6 Pro and Flash are open-weight models trained in a livestreamed RL run, rivaling GPT-5.6 and Claude Opus on coding benchmarks.
What is MiMo V2.6?
MiMo V2.6 is a pair of open-weight language models from Xiaomi, released as MiMo-V2.6-Pro-RL and MiMo-V2.6-Flash-RL. Both are sparse Mixture-of-Experts models trained with a single large-scale reinforcement learning run that Xiaomi livestreamed, reportedly costing around $3.5 million. The Pro model lands close to GPT-5.6 Sol and Claude Opus 5 on several agentic coding and tool-use benchmarks, while Flash offers a smaller, more deployable alternative with similar training but far fewer active parameters.
TL;DR
- MiMo-V2.6-Pro-RL is a 1.02 trillion parameter Mixture-of-Experts model with 42 billion active parameters, a 1 million token context window, and native text, image, video, and audio support.
- MiMo-V2.6-Flash-RL is the smaller sibling at 309 billion total parameters and 15 billion active parameters, built for cheaper deployment while keeping most of Pro’s agentic capability.
- Xiaomi trained both models in one mixed reinforcement learning run spanning coding, general agents, visual tasks, and cybersecurity instead of separate runs per domain, using asynchronous GRPO across batches of 1,568 prompts times 16 rollouts.
- On Terminal Bench 4.0, an unsaturated agentic benchmark, Pro scores 34.9 versus GPT-5.6 Sol’s 39.9 and Claude Opus 5’s 49.0, putting it within reach of frontier closed models on a genuinely hard test.
- The grading system behind the RL run compares multiple rollouts against each other rather than scoring pass/fail, using rubric-based synthesis and advantage redistribution to reward better, shorter solutions.
- Running Pro locally requires enterprise-grade hardware (roughly 8 H100-class GPUs for full precision), while Flash is realistic on a couple of high-end workstation-class GPU setups.
- Xiaomi released not just weights but also training code, plus a separate 9B distilled Qwen-based model aimed specifically at agentic coding on lighter hardware.
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
How was MiMo V2.6 trained?
Xiaomi built MiMo V2.6 around what it calls “You Only RL Once”: a single mixed reinforcement learning run covering coding, general agent tasks, visual reasoning, and cybersecurity, rather than training separate specialist models and merging them later. Tasks from different domains and different harnesses were mixed into the same training batches so skills learned in one area could transfer to tasks the model never saw during training.
The RL method is Group Relative Policy Optimization (GRPO), run fully asynchronously at a large scale: 1,568 prompts per step, each sampled 16 times, producing billions of tokens per update. That scale is part of why the run reportedly cost around $3.5 million, a figure Xiaomi made visible by livestreaming the training process and showing benchmark scores climb over time.
The more distinctive piece is the grading system. A simple pass/fail reward can’t tell the difference between a clean solution and a working-but-ugly one. Xiaomi’s approach, called Groupwise Agentic Grading, has an agentic grader compare multiple rollouts against each other within the same batch. One component builds task-specific rubrics from contrasting examples (Groupwise Reward Synthesis); another re-ranks passing trajectories and shifts reward toward the cleaner, shorter ones (Groupwise Advantage Redistribution). Because the grader judges the model’s own outputs against each other, Xiaomi describes this as a self-improvement loop, with extra safeguards (environment hardening, adversarial screening, verifier cross-checks) built in to stop the model from gaming the reward function.
After the RL stage, a separate distillation step called MOPD2 (Multi-Prefix Multi-Teacher On-Policy Distillation) blends the model’s own rollouts with prefix-conditioned examples from stronger teacher models and supervised demonstrations, aimed at improving performance on tasks that are hard to verify automatically.
How does MiMo V2.6 compare to GPT-5.6 and Claude Opus?
On several published benchmarks, MiMo-V2.6-Pro-RL sits close to GPT-5.6 Sol and Claude Opus 5, though not ahead of them across the board.
On DeepSWE v1.1, Pro scores 71.9 versus Claude Opus 5’s 74.0 and GPT-5.6 Sol’s 73.0. On AutomationBench v1.0.6, Pro actually leads at 53.1, ahead of Opus 5 (50.3) and GPT-5.6 Sol (45.8). On Terminal Bench 4.0, a benchmark that’s reportedly far from saturated, Pro scores 34.9, behind Opus 5’s 49.0 and GPT-5.6 Sol’s 39.9, but still a meaningful result for an open model. On Terminal Bench 2.1, an older and more saturated version of the same test, the gap nearly disappears: Pro hits 89.9, edging out both Opus 5 (89.1) and GPT-5.6 Sol (88.8).
Cybersecurity benchmarks show a different pattern. Pro dominates CyberGym (94.0) and MiMo Cyber Bench (80.2), categories where closed-model scores weren’t reported for comparison, but trails badly on ExploitBench (47.9 versus Opus 5’s 70.0 and GPT-5.6 Sol’s 78.5). This suggests the RL run pushed hard on some cybersecurity skills (like the kind measured in CyberGym) without matching frontier models on more adversarial exploit-generation tasks.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Flash, the smaller model, tracks close behind Pro on most benchmarks despite having roughly a third of the active parameters, which is the more interesting result for anyone trying to actually deploy this locally.
What hardware do you need to run MiMo V2.6?
This is where the “open weight” label needs a caveat. MiMo-V2.6-Pro-RL is a 1.02 trillion parameter model with 42 billion active parameters per forward pass. Running it at full precision realistically needs something in the range of 8 H100-class GPUs, plus additional VRAM overhead for the KV cache, which grows substantially at the model’s 1 million token context length. This isn’t a model anyone runs on a home rig; it’s built for enterprise inference clusters.
MiMo-V2.6-Flash-RL is far more approachable: 309 billion total parameters with 15 billion active. Xiaomi’s own deployment guides suggest running it with SGLang or vLLM across 4 to 8 GPUs depending on configuration, and it can be squeezed onto a pair of high-end workstation GPU setups at lower quantization. Both models support quantized formats, including FP8, on Hugging Face.
For anyone who wants agentic coding without a GPU cluster, Xiaomi also released a distilled version, a 9B-parameter model based on Qwen, specifically tuned for agentic coding tasks. That’s the more realistic option for individual developers experimenting locally.
Both full models are also available through hosted options: Xiaomi’s own MiMo API platform, AI Studio, MiMo Desktop, and OpenRouter, which sidesteps the hardware question entirely.
What does MiMo V2.6 actually do well?
In hands-on testing using an agentic coding harness, MiMo V2.6 Pro showed strong performance on multi-step, tool-using tasks: web research combined with landing page generation, live API integration (tracking the International Space Station’s position accurately), and using Blender to generate and animate 3D assets. It also handled procedural 3D scene generation, though not as convincingly as some competing open models in side-by-side testing.
One noted weakness was spatial reasoning in 3D asset generation, described as decent but not standout. Xiaomi also published experiments suggesting the model can assist with tasks like formal math proofs (using the Lean proof assistant), material science, and biochemistry problems, without any domain-specific reinforcement learning targeting those fields. That’s notable because it implies the coding and agentic RL training generalized into scientific reasoning tasks, rather than requiring a separate training run for each domain.
Is MiMo V2.6 worth using over closed models?
For teams already committed to open-weight infrastructure, MiMo-V2.6-Pro-RL is currently one of the strongest open options on agentic and coding benchmarks, and its cost-to-performance ratio makes it attractive for high-volume use cases where API costs from closed frontier models add up quickly. Flash is the more practical everyday choice given its lower parameter count and inference cost.
For teams without existing GPU infrastructure or without a need for full data control, GPT-5.6 Sol and Claude Opus 5 still lead on several harder benchmarks (Terminal Bench 4.0, ExploitBench), and using them via API avoids the hosting complexity entirely. The choice comes down to whether open weights, cost control, and the ability to fine-tune matter more than squeezing out the last few points of benchmark performance.
Frequently Asked Questions
What’s the difference between MiMo V2.6 Pro and Flash?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Pro is a 1.02 trillion parameter Mixture-of-Experts model with 42 billion active parameters, built for maximum capability. Flash is a smaller 309 billion parameter model with 15 billion active parameters, trading some benchmark performance for much lower inference cost and hardware requirements.
Can I run MiMo V2.6 on a single consumer GPU?
No. Even the smaller Flash model needs multiple high-end GPUs to run at reasonable precision, and Pro requires a multi-GPU enterprise setup (roughly 8 H100-class GPUs for full precision). Xiaomi’s distilled 9B Qwen-based variant is the realistic option for lower-end hardware.
Why did Xiaomi livestream the RL training run?
The livestream showed benchmark scores improving in real time as the reinforcement learning run progressed, offering unusual transparency into training cost (around $3.5 million) and methodology. It also served to demonstrate confidence in the “You Only RL Once” mixed-domain training approach.
Is MiMo V2.6 better than GPT-5.6 or Claude Opus?
It depends on the benchmark. MiMo-V2.6-Pro-RL matches or beats GPT-5.6 Sol and Claude Opus 5 on some tests (AutomationBench, Terminal Bench 2.1) but trails on others (Terminal Bench 4.0, ExploitBench). It’s competitive rather than uniformly superior, with a notably better cost-to-performance ratio.
What license is MiMo V2.6 released under?
The model weights are published under the MIT license on Hugging Face, along with the training code, making it usable and modifiable for both research and commercial purposes.
