Xiaomi has released MiMo-V2.6 as two huge open-weight multimodal agent models and a 9B distillation, making its reinforcement-learning recipe inspectable while showing how misleading active-parameter labels can be for local deployment. The release matters because it pushes reward-based agent training into public weights, but the flagship checkpoints remain distributed-serving artifacts rather than consumer-GPU downloads.
Key facts
- Xiaomi released MiMo-V2.6-Pro-RL, Flash-RL, and a 9B distill.
- Pro declares 1.02T total and 42B active parameters; Flash declares 309B total and 15B active parameters.
- Both headline models accept text, image, video, and audio and advertise one million tokens of context.
- Xiaomi's primary model cards declare MIT licensing; the inspected model trees did not expose a separate LICENSE file.
The easily missed news is what Xiaomi means by “RL.” It is not a new transformer architecture. The company describes a mixed post-training run across coding, general agents, visual tasks, and cybersecurity. For each step it says it uses 1,568 prompts and 16 candidate rollouts. Its Groupwise Reward Synthesis builds task rubrics from contrasting attempts, while Groupwise Advantage Redistribution ranks successful attempts and shifts reward toward better trajectories. In plain language, a coding agent that passes a test through a short, safe, understandable repair is meant to receive more credit than one that also passes after wandering through unrelated edits.
That is a worthwhile target. Many agent evaluations flatten all success into one bit: it passed or failed. Real software work is not like that. Two plumbers may stop the leak, but one may flood the kitchen first. Xiaomi is trying to make the learner prefer the first kind of solution. Its cards also describe MOPD2, a post-RL distillation scheme that mixes agent rollouts with teacher and supervised prefixes. The Pro technical report and model card are the authoritative descriptions; the effectiveness is still Xiaomi's own claim. Xiaomi calls the 9B model a “starting point for future research on agentic RL.”
The deployment story needs an equally clear translation. Pro has 70 transformer layers and 384 routed experts, eight selected per token; Flash has 48 layers and 256 experts, again with eight selected. Sparse routing reduces the computation per generated token, but a server still needs the complete bank of expert weights available. Xiaomi's official examples use eight-way tensor parallelism for Pro and four-way tensor parallelism for Flash. The Flash repository is about 178 GB and Pro about 573 GB. Those are disk footprints of the shipped files, not estimates from parameter count. A “15B active” model is therefore not a 15B-style local download.
The locally relevant release is instead the 9B model. It is not a compressed trillion-parameter MiMo: Xiaomi says it is Qwen3.5-9B supervised-fine-tuned on MiMo-generated coding, general-agent, visual-coding, and cybersecurity data. The official BF16 files total 18.8 GB across four shards. That alone exceeds 16 GB of VRAM before a runtime loads metadata, activation buffers, and a context cache. Xiaomi does not publish a minimum GPU-memory requirement. A third-party Q8 GGUF is 9.53 GB, which makes a 16 GB card plausible for short-context quantized use, but it does not prove that the official BF16 release fits there.
Xiaomi reports strong results on coding, general computer use, visual coding, and cyber evaluations, with Pro generally above Flash and Flash slightly above Pro on CyberGym. That is useful vendor evidence, not an independent leaderboard verdict. The strongest counterargument from the release discussion is that benchmark trajectories can still produce brittle or hard-to-maintain work, and “open weights” does not disclose the datasets or full training machinery needed to reproduce the result. The model-card interface also shows inconsistent parameter-size badges relative to Xiaomi's architecture summaries; readers should use the explicit architecture numbers, while treating the UI discrepancy as unexplained.
The important result is not that a trillion-scale model suddenly became a laptop model. It is that a large consumer company has published weights and a detailed claim about optimizing the quality of successful agent behavior. That makes GRPO, mixture-of-experts, and reward-based fine-tuning more than abstract ideas: they are now visible in a release that also illustrates the practical gap between active compute and total serving memory.
For users, that gap should change the first evaluation question from how many parameters are active to which checkpoint, quantization, context length, and serving topology can I actually operate?
Originally published on Ground Truth, where every claim is checked against the primary source.
