DeepSeek and Moonshot AI tapped GPT‑6 Astra distillation for Kimi K3.1 Pro and V4.1‑Flash 1017 after Anthropic crackdown
anthropic deepseek openai
| Source: Techmeme | Original article
DeepSeek and Moonshot AI used GPT‑6 Astra distillation to train their Kimi K3.1 Pro and V4.1‑Flash 1017 models after Anthropic’s crackdown.
Chinese AI start‑ups DeepSeek and Moonshot AI have turned to OpenAI’s latest flagship model, GPT‑6 Astra, as a core ingredient for their newest offerings. According to The Information, both firms relied heavily on outputs from Astra to distill and train their own models – DeepSeek’s Kimi K3.1 Pro and Moonshot’s V4.1‑Flash 1017 – after a recent “crackdown” by Anthropic limited access to other high‑end systems.
The move underscores how quickly advanced capabilities can be replicated through distillation, a process that extracts a larger model’s knowledge into a smaller, cheaper version. Astra, released in early September, has already attracted strong demand, and its adoption by Chinese developers highlights a growing market for “boot‑strapped” AI that can compete with home‑grown research without the same compute budget. For OpenAI, the surge in external distillation raises questions about the longevity of its competitive edge and the protection of its intellectual property.
Industry observers note that the development could intensify geopolitical and regulatory scrutiny. U.S. antitrust enforcers have recently probed Nvidia’s partnership with Hugging Face, and similar attention may soon focus on OpenAI’s licensing and export controls, especially as Chinese firms demonstrate the ability to repackage its most powerful model.
What to watch next: OpenAI’s response to the growing distillation ecosystem, potential policy adjustments around model licensing, and whether Anthropic or other competitors will tighten access to their own systems. The trajectory of DeepSeek’s and Moonshot’s products will also signal how effectively distillation can bridge the gap between cutting‑edge research and commercially viable AI in the near term.
Sources
Back to AIPULSEN