Kimi K3 dropped last week, and Nathan Lambert and Florian Brand used it as a forcing function to map the full state of open models in 2025. The conversation covers Xi Jinping's WAIC speech committing China to open source as a national strategy, Qwen's announcement that its next major model will be open-weight, and the sustained performance of GLM 5.2 as a fine-tuning base. The Chinese lab roster, including Qwen, DeepSeek, and MiniMax, gets a direct comparison against the US open-model ecosystem.
The most substantive section is the disagreement over how to measure the open-closed frontier gap. Lambert argues the relevant benchmarks are agentic coding and computer use tasks, where a gap of even a few months has real commercial consequences given the current market for tools like Claude Code and Codex. Brand pushes back on the instinct to assign a single number to the gap at all, noting that every benchmark provider surfaces different results. The distillation debate, prompted by a Ben Thompson piece, gets 10 minutes of direct treatment starting at 34:58.
The full episode runs to 48 minutes and includes a frontier tier list and explicit predictions. Lambert notes Kimi K3 weights are not yet public as of recording, expected July 27, and flags that post-training the model requires a full node of B300s just to load weights. The mechanics of that constraint, and what it means for who can actually extract value from large open releases, is the thread worth following into the transcript.
[READ ORIGINAL →]