The week everything accelerated
Following the Kimi K3 release, it feels like everything is accelerating. Geopolitics of US vs China. Economics of open vs closed models. Security at the frontier of AI. The Interconnects podcast with Florian Brand sits down to talk through what just changed, what it means, and what to expect before the end of the year.
Xi Jinping gave a speech at the World Artificial Intelligence Conference in Shanghai in which he directly committed to openness and open source as a strategy. It was not a detailed state-of-the-union address, but it was a clear signal that the Chinese state is willing to tolerate, and possibly underwrite, frontier-scale open-weight releases.
Alibaba said the next big Qwen model would be open weight, which is a big change of things. The Qwen team has historically been more selective about which weights they release, and a 2.4 trillion parameter Qwen 3.8 with open weights would be a meaningful expansion of the open ecosystem.
What Kimi K3 actually changes
Kimi K3 is a 2.8 trillion parameter mixture-of-experts model with 896 experts and 16 active per token. It supports a one-million-token context window, accepts visual input, and targets long-horizon coding, reasoning, and knowledge work. It ranks #3 on the Artificial Analysis Intelligence Index behind Claude Fable 5 and GPT-5.6 Sol Max, and #1 in Frontend Code Arena.
The system constraints are real. The weights are only the start. The API was saturated, the weights were pending, and third parties have to serve a 2.8 trillion parameter sparse model, which is a different operational class from anything most developers have run before. Lambert expects K3's rollout to take time.
GLM 5.2, by contrast, already had downloadable weights and support in vLLM, SGLang, and third-party providers, which is why it became the default "good enough" open model in the weeks before K3.
The distillation fight
Lambert and Ben Thompson agree that distillation works. They disagree about where it helps and how much training work it replaces. Reasoning traces and tool calls can seed supervised fine-tuning, teach response form, or bootstrap a data engine in a strong domain. Anthropic has alleged that DeepSeek, Moonshot, and MiniMax used millions of Claude exchanges for reasoning, coding, tool use, computer use, and grading.
Thompson argues that distillation grows more valuable as reinforcement learning grows, because frontier models can serve as teachers instead of a lab building every RL environment itself. He also proposes making model-training data collection fair use and banning terms that prohibit distillation for US companies.
Lambert's conclusion is more measured. Distillation may save months in one domain, but it does not replace problem generation, environments, reward design, and the RL system. The teacher does the easy part. The hard part is still on the student lab.
The frontier tier list
Lambert puts Kimi and Z.ai in his current frontier tier, with DeepSeek and Qwen close behind. Brand expects MiniMax to release a model above one trillion parameters and remain in that group. The frontier tier is now five or six labs, not two or three, and the open-weight tier list is no longer dominated by Llama derivatives.
The forward-looking piece is the Qwen 3.8 release, which will test whether the Chinese open ecosystem can sustain a frontier-tier release cadence. If Qwen 3.8 ships at the announced scale with open weights, the best open teacher available for distillation jumps an order of magnitude in a week, which is what changes the cost curve for anyone building on top of open models, more than any single leaderboard delta.
The week of WAIC was the moment when the open vs closed debate stopped being about the West and started being about everyone.
What the open-closed gap looks like at the end of 2026
Lambert's read on the conversation is that the gap is not closing toward zero, but it is no longer widening the way most observers expected. The closed labs still have structural advantages in training compute, data, and engineering depth, and the open ecosystem is still constrained by deployment complexity and licensing friction. But the open ecosystem is professionalizing fast, with multiple serious contenders at the frontier tier and a clear pipeline of releases through the rest of 2026.
The policy implications are real. If open-weight models at or near the frontier are available, then any domestic regulation of closed frontier systems becomes a comparative advantage for the open ecosystem. That is the structural reason the open vs closed question has become a US-China policy question as much as a technical one. The labs that build open models and the states that host their servers will have a different regulatory exposure than the labs and states that do not.
The week's news made the trend harder to dismiss. The next release cycle will determine whether it is now a structural feature of the AI industry or a temporary phase that the closed frontier will eventually re-establish dominance over. The bet, from the Interconnects read of the week, is that it is a structural feature.