📢 Flash
Why Did DeepSeek’s New Version Suddenly Get So Strong? Foreign Media Exposes Astra Distillation Secrets—Liang Wenfeng Yields to Competitive Pressure
Beating AI News reports that DeepSeek’s newly released model shows significant performance improvements—potentially powered by OpenAI’s GPT-6 Astra. According to *The Information*, citing multiple insiders directly involved in the training process, DeepSeek’s V4.1-Flash-1017 version—launched on October 17—and Moonshot’s concurrently released Kimi K3.1 Pro both leveraged Astra’s outputs at scale for knowledge distillation during training.
Notably, DeepSeek’s latest model excels in browser operation capabilities: on Browser Use—a benchmark covering 180 real-world web tasks—its score rose from 70.9 (previous version) to 75.0, nearing Astra’s own 77.9. Yet its per-task inference cost stands at just $0.22, dramatically lower than Astra’s $10.77—roughly 49x cheaper.
Previously, Chinese AI firms primarily used Anthropic’s Claude series as their main distillation target. However, over the past two months, Anthropic has steadily strengthened its model safeguards, making it significantly harder to extract reasoning traces—prompting domestic labs to pivot their distillation efforts toward Astra.
Interestingly, DeepSeek co-founder Wenfeng Liang had initially opposed incorporating distillation into the V4 series’ development. But after observing rapid iteration by domestic rivals—including Zhipu AI’s GLM models via distillation—he ultimately yielded to competitive pressure.
Notably, DeepSeek’s latest model excels in browser operation capabilities: on Browser Use—a benchmark covering 180 real-world web tasks—its score rose from 70.9 (previous version) to 75.0, nearing Astra’s own 77.9. Yet its per-task inference cost stands at just $0.22, dramatically lower than Astra’s $10.77—roughly 49x cheaper.
Previously, Chinese AI firms primarily used Anthropic’s Claude series as their main distillation target. However, over the past two months, Anthropic has steadily strengthened its model safeguards, making it significantly harder to extract reasoning traces—prompting domestic labs to pivot their distillation efforts toward Astra.
Interestingly, DeepSeek co-founder Wenfeng Liang had initially opposed incorporating distillation into the V4 series’ development. But after observing rapid iteration by domestic rivals—including Zhipu AI’s GLM models via distillation—he ultimately yielded to competitive pressure.
