Rumors: Chinese open-weights models advanced by extracting reasoning traces from Claude Code and Codex

jxmnop · x · 2026-08-10

An AI researcher shared interesting industry rumors regarding the distillation techniques behind open-weights models like Kimi, Qwen, and Minimax. Good distillation heavily relies on reasoning traces, which are normally hidden from users.

Speculatively, Chinese labs figured out how to reverse-engineer reasoning from Claude Code and Codex early on, allowing them to collect large amounts of long-horizon data with reasoning traces included. This breakthrough allegedly fueled the recent wave of powerful open-source models. The author notes that reasoning extractability remains a massive uncertainty for the future of open models, aligning with Anthropic's warnings about distillation attacks.

Original post →

More from Models

Models channel →