Open Models: Kimi K3, Qwen 3.8, Xi’s WAIC Speech, Distillation, The Open-Closed Gap, and What’s Next

This is a 49-minute Interconnects podcast conversation between Nathan Lambert and Florian Brand, published July 22, 2026. They discuss Chinese and US open models, using Kimi K3 and GLM 5.2, distillation, access to frontier systems, and likely releases through the end of 2026.

Moonshot launched Kimi K3 as a 2.8-trillion-parameter model and promised full weights for July 27. Xi Jinping's WAIC keynote called for open source, openness, collaboration, and sharing. Qwen said Qwen3.8 would also be open weight.

Lambert had expected capital-intensive frontier training to widen the gap between closed and open models. Chinese open models keep closing it.

The gap changes by benchmark, workload, and serving setup. K3 can score near the frontier while remaining hard to deploy. Its 2.8 trillion parameters put it in a different systems class from the smaller open models most developers can run or tune. The weights are only the start.

Brand found K3 useful for exploratory research. In one experiment, it used tools to search Reddit for signals that might lead the stock market by one or two months. Its experimental code was simple and readable.

He still ran Codex afterward. That pass found specialized misses and added edge cases. K3 joined his workflow without replacing it.

K3 was also slower in wall-clock time, and launch demand strained the service. Kimi's paid plans ranged from $19 to $199 a month. Higher tiers received more Kimi Code credits and longer context, but overload and request-limit failures still occurred. Moonshot paused new subscriptions while adding capacity.

GLM 5.2 remained fast, dependable, and good enough for many jobs. A model already integrated into a workflow can beat a stronger one that is slow or hard to access.

Lambert expects K3's rollout to take time because the API was saturated, the weights were pending, and third parties must serve a 2.8-trillion-parameter sparse model. GLM 5.2 already had downloadable weights and support in vLLM, SGLang, and third-party providers.

Brand expects providers to move faster. They now prepare before major releases, and the serving ecosystem is more professional than it was a year ago. Both expect K3 to spread; they disagree on the speed.

Lambert proposes that Chinese labs turn money into compute, data, talent, and capability more efficiently. Possible reasons include lower compensation, subsidies, technically trained teams, and the narrower task of catching a visible frontier instead of finding it first. A Kimi engineer told him that catching up is cheaper because the target is known.

Brand describes teams of roughly 200 to 300 mostly young researchers focused on one model. These labs also serve fewer paying users than ChatGPT or Anthropic, leaving more compute and attention for training.

Compute access has grown. Lambert includes advanced NVIDIA hardware routed around export restrictions. Brand points to domestic accelerators. Meituan says LongCat-2.0's training and deployment ran on AI ASIC superpods across more than 35 trillion pretraining tokens and millions of accelerator-days. Domestic chips are especially useful for inference-heavy reinforcement learning.

Lambert says the cause is still unclear. The result is clear enough to require an explanation: the expected capital wall has not pushed Chinese open models further behind.

Commercial work can distract a research lab. Safe Superintelligence designed its organization to avoid management overhead and product cycles, giving Lambert a test case for concentrated research.

Labs can also buy more of their training stack. Scale says nearly half of its new data-training projects involve RL environments. Mercor sells expert datasets and simulated worlds with tools, tasks, and verifiers.

Brand thinks later buyers may get the same environments more cheaply and catch up without building every input first. He says public information cannot show how large this effect is.

Qwen combines a broad model family with Alibaba Cloud and a large developer ecosystem. Lambert considers its smaller models especially important because researchers can use them directly.

DeepSeek exposes a large V4 Pro model and a smaller V4 Flash model through preview APIs. The hosts keep it near the frontier but find its release behavior hard to read.

MiniMax and Zhipu are public companies, adding investor pressure to product decisions. MiniMax uses a restrictive license. Kimi had promised weights without naming a license in its launch article.

Ant Ling spans general, reasoning, and multimodal models. Meituan's LongCat is tied to internal enterprise use and domestic hardware. These providers have less global developer reach than Qwen, DeepSeek, or Kimi, but increase the number of labs capable of strong releases.

Thinking Machines released Inkling, a 975-billion-parameter model with 41 billion active parameters, and previewed a smaller version. Inkling is built as a customizable base that users can fine-tune through Tinker.

Poolside is releasing small sparse coding models faster. NVIDIA remains the most consistent US participant through models, data, and tools. Google has Gemma, OpenAI has gpt-oss, and Meta has Llama.

Lambert and Brand see an opening in useful sizes, reliable fine-tuning, regular releases, and developer familiarity. Company and investor pressure still pushes labs toward expensive frontier claims.

Near-frontier capability is getting cheaper and easier to access. Frontier systems are becoming more controlled. Anthropic limited Claude Mythos Preview after stronger cybersecurity results. OpenAI distributes GPT-Rosalind for life-sciences work through trusted access for qualified customers.

Hugging Face's July security incident shows why local access can matter. Its responders analyzed more than 17,000 logged events from an intrusion involving tens of thousands of autonomous-agent actions. Hosted frontier APIs blocked real attack commands, exploit payloads, and command-and-control artifacts because they could not distinguish incident response from attack. Hugging Face ran GLM 5.2 on its own infrastructure and kept sensitive data inside its environment.

Hugging Face said this was not an argument against hosted safeguards. Its lesson was that defenders need a vetted local model when legitimate work collides with API controls.

Lambert applies the case to contemplated US restrictions on Chinese open models, including export-list action, federal advisories, hosting requirements, and supply-chain powers. These were proposals, not an enacted blanket ban. He argues that crude restrictions could remove tools from US defenders and developers without stopping the capability from spreading elsewhere.

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 says DeepSeek, Moonshot, and MiniMax used millions of Claude exchanges for reasoning, coding, tool use, computer use, and grading.

Thompson says 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 supports that legal proposal but rejects the training argument. Large RL runs still need prompts, environments, millions of rollouts, rewards, and fast grading. Calling a slow frontier API as a grader at that scale could be expensive and throughput-limited. A tailored local grader may work better.

Brand separates the stages. SFT teaches manners, formatting, and initial behavior. RL pushes stronger capabilities. The benefit of SFT data from the newest frontier model over an older open model remains unknown.

OpenThoughts3 found that question mix, filtering, and multiple answers change the student result. OpenThoughts-Agent found that the strongest benchmark model was not automatically the best teacher. Lambert's conclusion is that distillation may save months in one domain but does not replace problem generation, environments, reward design, and the RL system.

K3 fulfilled Brand's earlier prediction that an open model would pass two trillion parameters. He now expects slower growth. A model just above three trillion is plausible; five or ten trillion would surprise him.

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 US wild cards are NVIDIA, Thinking Machines, and Reflection. Lambert thinks one could enter the top cluster by late November. Brand says their first advantage may be as bases for domain fine-tuning rather than ready-made frontier agents.

Tencent is Brand's other wild card. Hy3 has 295 billion total parameters, 21 billion active parameters, and an Apache 2.0 license, replacing the earlier Hy3 Preview community license. He expects a larger release by year end.

Both expect surprises and treat the tier list as provisional.