Xiaomi MiMo-V2.6-Pro Tested: Does the 1T-Param Open Model Hold Up?
Xiaomi's 1.02T-parameter MoE model MiMo-V2.6-Pro tested on coding, design, and vision tasks, with benchmark comparisons to GPT and Claude.
What is Xiaomi MiMo-V2.6-Pro?
MiMo-V2.6-Pro is Xiaomi’s flagship open-weight AI model, a 1.02 trillion parameter mixture-of-experts (MoE) system that activates only 42 billion parameters per forward pass. It handles text, image, video, and audio in one model, supports a 1 million token context window, and is trained with a single mixed reinforcement learning run spanning coding, general agent tasks, visual reasoning, and cybersecurity instead of separate training runs per skill. Hands-on tests show it competing with Claude Opus 5 and GPT-5.6 on real coding and agent benchmarks, though it’s slower and more token-hungry on some tasks.
TL;DR
- Xiaomi’s MiMo-V2.6-Pro-RL is a sparse MoE model with 1.02 trillion total parameters and 42 billion active parameters per token, released as open weights on Hugging Face.
- The model uses one mixed reinforcement learning run across coding, agentic, visual, and cybersecurity tasks rather than training each domain separately, which Xiaomi says lets skills transfer across domains.
- On benchmarks like DeepSWE v1.1 (71.9) and JobBench (62.0), it lands close to Claude Opus 5 and GPT-5.6, though it trails both on harder agent benchmarks like Terminal Bench 4.0 and cybersecurity-heavy tests like ExploitBench.
- A live-streamed RL training run reportedly took MiMo-V2.6-Pro through 30 RL steps over about 750,000 trajectories in under 6 days, lifting its DeepSWE v1.1 score from 58.4 to 72.57 for roughly $2.62 million in compute.
- In a real-world coding test on a broken ferry occupancy dashboard, the model correctly diagnosed and fixed a subtle data-aggregation bug without being told what was wrong, but it took over 25 minutes and used a heavy number of tokens compared to faster models like Grok 4.7.
- On a creative front-end design brief for a coffee shop landing page, the model produced a page with genuine personality (custom copy, layout choices, working navigation) rather than a generic template.
- On a vision-to-game-logic task, the model correctly identified objects in a complex image and mapped them to interactive game mechanics, but the resulting playable prototype shipped with a keyboard input bug it couldn’t fully resolve after one fix attempt.
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.
How does MiMo-V2.6-Pro’s architecture work?
MiMo-V2.6-Pro is a sparse mixture-of-experts transformer with 384 routed experts, of which 8 are activated per token. The backbone runs 70 layers total, split between 60 sliding-window attention (SWA) layers and 10 global attention (GA) layers, with a hidden size of 6144. The very first block uses global attention with a dense feed-forward layer, and the rest interleave local and global attention using MoE feed-forward layers without shared experts.
For multimodal input, Xiaomi built a dedicated 681-million-parameter vision encoder (MiMo ViT, 28 layers) and separate audio components: a 308-million-parameter audio tokenizer with 20 RVQ codebooks and a 127-million-parameter audio patch encoder. A 5-layer multi-token prediction module acts as a speculative decoder, predicting several tokens ahead per forward pass to speed up inference. The model supports a 1 million token context window, aimed at long agent sessions, large codebases, and multi-turn tool use.
What makes the “one RL run” training approach different?
Most large models train reasoning, coding, and agentic skills in separate reinforcement learning pipelines, then merge the results. MiMo-V2.6-Pro instead mixes coding, general agent, visual, and cybersecurity tasks into the same RL batch, using Group Relative Policy Optimization (GRPO) at scale, 1,568 prompts times 16 rollouts per step, processing billions of tokens per update.
Xiaomi also changed how it grades rollouts. Instead of simple pass/fail rewards, an agentic grader ranks passing solutions against each other within each group. This is done through two mechanisms: Groupwise Reward Synthesis, which builds task-specific rubrics from contrasting rollouts, and Groupwise Advantage Redistribution, which shifts training signal toward higher-quality solutions among the ones that already passed. The idea is that this closes a self-improvement loop that also nudges the model toward shorter, cheaper solution paths, since raw pass/fail rewards give no incentive to be efficient.
Xiaomi streamed part of this training process publicly. According to figures cited from that stream, the Pro model ran 30 RL steps over about 750,000 trajectories in under 6 days, at a compute cost of roughly $2.62 million, and its DeepSWE v1.1 score rose from 58.4 to 72.57, a 14-point jump that reportedly generalized to tasks outside the training set.
How does it perform on coding tasks in practice?
In a hands-on test, MiMo-V2.6-Pro was pointed at a full-stack ferry occupancy dashboard app running locally, with a real bug: the booking system counted each booking record as one passenger regardless of group size, so a 30-person booking showed up as occupancy of one. This under-reported real occupancy and created an overbooking risk. The model was given no hint about the bug’s location or nature, only a general goal to find and fix it.
It succeeded, correctly identifying that the booking endpoint counted rows instead of summed passengers, and fixed the aggregation logic so the dashboard reflected real seat counts after the fix. However, the process took over 25 minutes and consumed a large number of tokens for what was described as a fairly contained bug, a noticeably slower and pricier result than a comparable fix attempted on Grok 4.7 the same day. On raw benchmark numbers, MiMo-V2.6-Pro scored 71.9 on DeepSWE v1.1 and 63.2 on Xiaomi’s own MiMo Code Bench, both close to but slightly behind Claude Opus 5 (74.0 and 68.6 respectively) and ahead of Claude Fable 5 (70.0) on DeepSWE.
Is MiMo-V2.6-Pro good at design and front-end work?
Yes, based on a real-world creative brief test. The model was given a detailed prompt for a Melbourne coffee shop landing page with instructions to avoid generic template defaults and lean into bold colors and an editorial layout. It produced a working page with custom copy (including an invented backstory about the shop getting evicted from a garage), a distinct visual style, and functioning navigation between menu, story, and contact sections. It also started a local Python server on its own and returned a working URL without being explicitly walked through each step, and created a reusable “skill” from the task without being asked. This lines up with Xiaomi’s own framing of the model having a “taste for front-end” work, and the visual coding benchmark score (72.3 on MiMo VisualCoding) sits close to GPT-5.6 Sol (73.4) and above Claude Fable 5 (69.1).
How well does it handle vision and multimodal tasks?
The vision test involved giving the model a complex image and asking it to translate it into a playable 3D game concept, then generate an actual HTML prototype. The model correctly identified numerous objects in the scene (balls, mice toys, cat towers, a donut-shaped bed, among others) and mapped each to a specific interactive game mechanic rather than producing a flat description. It then generated a single-file HTML prototype with working physics, including a functioning “pounce” action triggered by the space bar.
The prototype shipped with a bug, though: the WASD movement keys didn’t register in the browser (confirmed across Chrome and Edge), and a developer console check showed no obvious JavaScript errors. When told about the bug, the model attempted a fix, but the keys remained non-functional afterward. This suggests the model’s image-to-logic reasoning is strong, but its ability to fully debug front-end input handling in one pass has limits.
How does it compare to GPT-5.6 and Claude Opus 5 on benchmarks?
Across the published evaluation table, MiMo-V2.6-Pro is competitive but not dominant. It leads on JobBench-adjacent agent tasks like AutomationBench v1.0.6 (53.1 versus Claude Opus 5’s 50.3 and GPT-5.6 Sol’s 45.8) and posts strong cybersecurity numbers on CyberGym (94.0) and MiMo Cyber Bench (80.2), categories where comparison figures for GPT-5.6 and Claude Opus 5 weren’t available. It falls behind on harder agent evaluations like Terminal Bench 4.0 (34.9 versus Opus 5’s 49.0) and lags noticeably on cybersecurity exploit benchmarks like ExploitBench (47.9 versus Opus 5’s 70.0 and GPT-5.6 Sol’s 78.5). Overall, for a model of this size going head-to-head with frontier closed models on several categories, the results are respectable without being a clean sweep.
Frequently Asked Questions
How many parameters does MiMo-V2.6-Pro have?
It has 1.02 trillion total parameters in a mixture-of-experts architecture, with only 42 billion parameters activated per token during inference, out of 384 total routed experts (8 activated at a time).
Is MiMo-V2.6-Pro open source?
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
Yes. It’s released as open weights on Hugging Face and ModelScope under Xiaomi’s MiMo project, alongside a smaller MiMo-V2.6-Flash-RL variant, with deployment support through SGLang and vLLM.
What context window does MiMo-V2.6-Pro support?
It supports up to 1 million tokens of context, intended for long codebases, extended tool-use traces, and multi-session agent workflows.
How does MiMo-V2.6-Pro compare to Claude Opus 5 and GPT-5.6?
It performs in a similar range on several coding and agent benchmarks, sometimes ahead (like AutomationBench and JobBench) and sometimes behind (like Terminal Bench 4.0 and cybersecurity exploit tasks), making it a strong open alternative rather than an outright leader.
What are the practical weaknesses seen in hands-on testing?
The main issues observed were slower response times and higher token usage on a moderately complex coding bug compared to other current models, and an unresolved keyboard input bug in a generated 3D game prototype even after a follow-up fix attempt.
