Six months of open-source LLMs: Xiaomi MiMo, DeepSeek and Chinese models on Hugging Face

Competition among Chinese open-source large models has accelerated over the past six months. On September 22, Xiaomi ended a six-day "alchemy livestream" that ran reinforcement learning training on camera. The final bill was 3.5 million US dollars, about 23.44 million yuan.

Results were published alongside the stream. MiMo-V2.6-Pro has 1.02 trillion parameters and 42 billion active. It scored 46 on the Artificial Analysis Intelligence Index, first among open-source models. MiMo lead Luo Fuli said this was one of the largest single reinforcement learning runs any open-source team has carried out. She noted that the research and engineering challenges exceeded those of DeepSeek-R1, which she had worked on.

Out-of-sample generalization is the key takeaway. On DeepSWE v1.1, a benchmark for coding agents, Pro rose from 58.4 to 72.57 points, and Flash from 48.7 to 65.68. After only 30 steps, the models improved and transferred their abilities to unseen software engineering tasks.

Prices did not rise with capabilities. MiMo-V2.6 keeps the previous generation's API pricing: Pro costs about 3 yuan per million input tokens and 6 yuan for output, Flash about 1 yuan and 2 yuan. Artificial Analysis estimates that MiMo-V2.6-Pro completes a task for an average of 0.13 US dollars. The closed-source Grok 4.7, released in the same period, also scores 46 on the index, but its high-end version costs an average of 3.74 US dollars per task.

The open-source release list is growing. Xiaomi released model weights, the technical report, more than 7,000 RL task environments, an end-to-end training framework, and a freely combinable mini-harness. It also released a distilled model built on a Qwen3.5-9B base, supervised-fine-tuned on 77.4B tokens of data generated by MiMo. A former Hugging Face researcher highlighted the speed: the final RL training started less than a week before the model and framework were delivered.

DeepSeek uploaded DeepSeek-V4-Flash-Vision-Exp, the first multimodal model in the V4 series, to Hugging Face on August 31 under the MIT license. The release includes model files, the tokenizer, a reference prompt encoding implementation, and a minimal PyTorch inference implementation. Driven by both environment and ecosystem, Chinese models are becoming more visible in the open-source community.