Open models trail the frontier by three months, and Lambert argues distillation does not explain it

Nathan Lambert has published a recording of the open models lecture he gave on October 22 in the AI Safety course at Harvard. His main thesis: open models trail closed ones by only a few months, even though American labs have roughly ten times more compute according to the sources he cites.

Chinese open models keep pace with a tenth of the compute

Lambert previously co-led the Olmo models at Ai2 and now directs the nonprofit Trillium Labs. The lecture runs 48 minutes, and he says he wants to update it roughly every six months. This piece relies on his public slides, which carry the figures below.

On the Artificial Analysis index open models trail the frontier by about three months, and on Epoch AI's software engineering index by about five. American open models, Lambert says, trail by a further three to four months. Among the strongest recent models he lists Xiaomi MiMo-V2.6-Pro, with a trillion parameters and a score of 46, and DeepSeek V4.1 Flash with 45.

Compute sits on the other side of the ledger. According to the estimates Lambert cites, leading American labs have roughly ten times the compute of their Chinese counterparts. The IFP institute puts the average advantage of the top five US labs at 16 times by the end of 2026.

Lambert calls distillation a convenient explanation, not the main engine

American labs often explain Chinese progress by distillation, mainly meaning extracting hidden reasoning from their APIs and training on it. Lambert acknowledges that such teams are common in China. Yet public research, he argues, does not show that borrowed reasoning traces produce an easy jump in performance. Reinforcement learning plays a growing role. Radar has previously covered his critique of how US agencies use the word distillation.

As evidence he points to Xiaomi, which streamed the MiMo-V2.6 training live on a public dashboard. The Pro version ran 30 RL steps in 5 days and 7 hours, processed 75 billion tokens and cost $2.62 million. The smaller Flash came in at $854 thousand. Lambert's conclusion: if distillation were banned, Chinese labs would remain strong competitors, because fast following costs far less compute than pushing the frontier.

For companies deploying open models, the more practical part of the lecture is licensing. DeepSeek, Qwen, gpt-oss and Gemma 4 ship under MIT or Apache 2.0. Kimi K3 and Muse Spark Open, however, require a separate agreement above $20 million in revenue for inference services, and GLM-5.3 requires a security review above $10 billion. Lambert notes that most of these terms have not been tested in court.

A lecture for a safety course that mostly makes the case for openness

Lambert is an advocate of open models and makes no secret of it. To a safety course audience he argues that banning open models would mainly hobble economic activity, while bad actors would get access anyway. He concedes that open weights clearly accelerate some harms, such as non-consensual deepfakes. The cyber incidents on his slides, though, mostly involve closed models, which conveniently supports his argument.

He admits the weak spots himself. He calls Hugging Face downloads an incredibly imperfect proxy. There is very little public data on the impact of distillation, he says, so his own position is closer to an informed estimate than proof. He also grants that Chinese labs may care a bit more about public benchmarks and that American models tend to be more well rounded in the long tail.

In six months we will know whether the gap holds once APIs are locked down

Lambert himself says the labs are only starting to secure their APIs against distillation. If the gap widens after that, the labs were right. If it stays around three months, his reading holds: research and in-house RL are what decide the race.

The second signal is demand for American open models. According to Lambert, US-made models of similar performance get two to five times more downloads than Chinese ones. If American labs close the performance gap, they can offer companies a model without geopolitical question marks. Until then, Chinese models lead on performance, downloads and research citations.

Lilith's verdict

Anyone blaming a rival's progress on copying should also explain why Xiaomi livestreamed its model's training, price tag of $2.62 million included. Once the labs lock their APIs and the gap still sits at three months, America is left with the one advantage nobody can download: gigawatts in data centers.