This week, what looked like a simple question on Reddit — "Where's the RL version of Xiaomi MiMo V2.6 9B?" — exposed a new phase in China's open-source LLM race: from competing on speed to competing on completeness.
What this is
Xiaomi joined the open-source battlefield this year with the MiMo series. A user in LocalLLaMA (the developer community for locally deployed open-source LLMs) tried to download the "Distill RL" version of MiMo V2.6 9B — lightweight weights produced through distillation (a small model learning to mimic a larger one) and further refined through reinforcement learning (additional training under reward signals) — and discovered that the official repository only offered Base (foundation) and Instruct (instruction-tuned) versions, with the RL weights never released. Several regular contributors to the open-source circle confirmed in follow-up comments: there is currently no public channel.
Behind a missing download link lies an old debate in the open-source ecosystem: when a model is only half-released, can developers really trust it?
Industry view
We have noticed that China's leading LLM companies have significantly ramped up their open-source activity this year: DeepSeek, Alibaba's Qwen, ByteDance's Seed, Zhipu's GLM, and now Xiaomi's MiMo. But the industry is starting to worry about "open-source inflation" — the more models released, the harder it becomes to verify their completeness.
Pro side: RL weights involve training data and reward strategies, so there is reasonable justification for withholding them; most API callers don't need RL checkpoints — they're only useful for teams doing fine-tuning (continuing to train the model on their own data).
The con side is sharper. An engineer who has tracked open-source licensing for years told us: "Releasing the Base model but not the RL version means there's always a wall between what users see on the API and what they can reproduce locally. It's commercial protection in the short term, but it costs trust in the long run." Others point out that half-open-source releases end up satisfying no one — not as thorough as Meta's Llama, and not as clean as OpenAI's or Anthropic's outright closure.
Impact on regular people
For most enterprise IT teams, differences in open-source model "completeness" don't affect production yet — API calls remain the default.
For individual careers, this story is largely irrelevant to day-to-day work — unless your company is building local models from scratch, you don't even need to know what MiMo is.
For the consumer market, the indirect signal: the more mature the open-source ecosystem, the lower the cost of local inference, raising the possibility that small and medium businesses will eventually run dedicated models on their own servers — but this is a slow-moving variable over the next 3-5 years.