DeepSeek and Kimi use Astra output for model training, resulting in improved performance.

ME News reports that, on October 29 (UTC+8), according to The Information and multiple directly informed sources, DeepSeek’s updated V4.1-Flash-1017 released on October 17 and Moonshot AI’s newly released Kimi K3.1 Pro both extensively utilized outputs from OpenAI’s GPT-6 Astra for distillation during training. DeepSeek’s new model showed improved performance in browser operation tasks. In a test of 180 real-world web tasks on Browser Use, its score rose from 70.9 in the previous version to 75.0, nearing Astra’s 77.9; the cost per model invocation is $0.22, compared to Astra’s $10.77—a nearly 49-fold difference. Additional insights indicate that previously, Claude was the primary target for distillation by Chinese AI companies, but Anthropic has tightened its defenses over the past two months, making it more difficult to extract model reasoning processes, prompting domestic labs to shift focus to Astra. DeepSeek’s founder, Liang Wenhong, originally resisted relying on distillation in the development of V4, but ultimately yielded to competitive pressure after observing domestic competitors like Zhipu’s GLM making progress through distillation. (Source: ChainCatcher)