Open models recap: more on Kimi K3, Qwen 3.8, Xi's WAIC speech, distillation, the open-closed gap, and what's next
Interconnects · Nathan Lambert · 2026-07-22
Interconnects podcast hosts Nathan Lambert and Florian Brand argue that Chinese AI labs like Moonshot AI (Kimi K3) and Zhipu (GLM 5.2) are closing the frontier gap through capital efficiency and RL-centric training rather than distillation from closed models, while warning that banning these open-weight models would harm U.S. cybersecurity defenders more than attackers.
Appears in
Extraction
Topics: open-weight-modelschinese-ai-labsdistillationreinforcement-learningai-geopolitics
Claims
- Kimi K3 performs at approximately GPT-5.4–5.5 level on coding tasks and demonstrates strong agentic research capabilities, suggesting Chinese frontier open models are near-competitive with leading closed models.
- Distillation from closed model APIs has limited impact on Chinese model performance because RL training—where real capability gains occur—requires millions of rollouts that would be prohibitively expensive and slow to run through third-party APIs.
- Ben Thompson's Stratechery argument that distillation becomes more impactful as RL scales is factually incorrect and potentially misleading to policymakers who may use it to justify restrictive action.
- Chinese labs' cost and capital efficiency advantages stem from focused engineering teams with no distracting side projects, lower operational costs, and increasing access to domestic compute including Huawei Ascend chips.
- Banning Chinese open-weight models from U.S. companies would harm American defenders in cybersecurity, as Hugging Face could only analyze the OpenAI breach using a Chinese open model because U.S. frontier models had guardrails blocking the analysis.
- The open-weight model ecosystem has professionalized significantly, with Chinese labs now releasing strong models on 1–2 month cycles that match closed lab iteration speeds.
Key quotes
If I had a magical API that gave me reasoning traces from Claude/Gemini, I actually don't know if me like fine-tuning an OLMo model on that would make OLMo smarter. It's one of the most wild unanswered research questions.
These are not just benchmaxxed distilled IP theft models. These are like genuinely good models that people are comparing to on their internal training benchmarks.
If the rest of the world has access to these open models and we ban them for the companies in the US to use… it's just like a growing disparity between US companies ability to defend and the attackers all over the world in terms of cyber.