Open weights are only 3-5 months behind closed — and that time gap is the entire safety margin
Kimi K3 compresses the open-versus-closed gap from a contested 6-9 months to 3-5 months, and that gap is the only risk buffer there is — heavy-handed regulation of open weights would only delay the inevitable while voiding the buffer itself.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
The gap has been compressed to 3-5 months
K3 is a 2.8T-parameter MoE, released July 16 with weights promised for July 27. The author pins the whole piece's key conclusion on one number: the performance gap between open and closed, or between the US and China, has narrowed from the contested 6-9 months to 3-5 months. He adds a robustness note — even if Moonshot ultimately doesn't deliver on open weights, as long as a comparably strong closed model exists in China, most of the inferences only land in an intermediate zone, and the direction is unchanged. The leaderboard evidence: second overall on Vals AI, third on the Artificial Analysis intelligence index (behind only Claude Fable and GPT-5.6 Sol Max, and cheaper), first on Frontend Code Arena. This is the closest an open model has come to the frontier since DeepSeek R1.
— Nathan LambertThe distillation narrative dies in this episode
The author states the conclusion outright: if adversarial distillation contributed to K3 at all, it was only to a small degree. The reason is that R1 and K3 are different in kind — R1 was a Chinese lab sprinting to pivot to reasoning models first, whereas K3 is scaling along known axes: data, algorithms, architecture, tooling, environments — the same set of problems people at OpenAI and Anthropic are working on. Another layer of evidence comes from his trip to China: the culture of the Kimi team was one of the few among the companies he visited that you could feel at a glance, with strong individual execution and freedom of expression even under GPU constraints; having seen it, the result was unsurprising to him. He predicts the distillation debate and the associated pressure will continue, but that observers attributing Chinese models to IP theft will be embarrassed by reality.
— Nathan LambertThe WAIC speech is a statement of risk tolerance
Previously there was almost no national-level policy statement from China on open-source AI; the author says that to his knowledge no senior leader had commented publicly. That changed this week: in his keynote at the World Artificial Intelligence Conference (WAIC), Xi Jinping very directly bet the future of China's AI ecosystem on open source and global diffusion. The author's reading is that this commitment, landing in the same week as the strongest open-weights model in history, amounts to an implicit statement of China's risk tolerance for open weights. The simplest explanation: the Chinese government is tracking model risk closely, at a technical granularity that may be finer than America's ‘vibes-based regulation’, and will act if it measures risk — and it currently judges frontier models to pose no substantial risk. At the same time, economic policymakers think getting AI used first is a good thing, scaling distribution before monetising, a path consistent with cars, solar and advanced manufacturing.
— Nathan LambertOpen source is the frontier labs' economic soft spot
Dean Ball (personally pro-open-source, but now at OpenAI) says open-weights models are essentially decelerationist, and the author says he's right. The transmission chain is concrete: strong open models sharply compress the gross margin space of closed labs — first, the labs have less profit to reinvest in the next generation of models; second, the market marks down these companies' terminal values, discounting future funding rounds. Together these delay the arrival of the most transformative models. But the author doesn't think this is enough to knock OpenAI and Anthropic out of the ranks of the world's most valuable companies, and he thinks it's a net good for society: open source is accelerationist on the diffusion side, lowering the entry price for a given performance tier of intelligence and encouraging customisation — it's just that this curve starts much more slowly. His phrasing: an exponent that starts slower but may be larger. The risk is that closed capabilities run so far ahead that customisation loses its meaning.
— Nathan LambertWhere the 2.5× efficiency comes from: three changes
Kimi's release blog gives checkable technical handles: K3 builds on two architectural updates, Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), improving information flow along sequence length and along model depth respectively; it also pushes MoE sparsity up to 16 of 896 experts actually activated, paired with the Stable LatentMoE framework. Together with refinements to the training and data recipe, this yields roughly a 2.5× overall scaling-efficiency gain over K2. The author also traces the diffusion path of this innovation: KDA comes from the Kimi Linear paper and shares an origin with the Gated DeltaNet he used when building Olmo Hybrid at Ai2; Gated Delta Networks were only proposed at the end of 2024, built on ideas from Mamba, and by mid-2026 they are in frontier models — Qwen's latest model has also switched to a related architecture.
— Nathan LambertCapital efficiency: catching up is cheaper than inventing
The author judges Chinese labs to be markedly more capital efficient, and in a world where intelligence is proportional to effective capital (buying compute, data, talent), that may be the single biggest advantage an AI industry can have. He lists several factors he's heard but can't verify: Chinese researchers are paid less yet are better at solving LLM research puzzles; a data industry is forming (budgets still far below Anthropic's billion-dollar scale); meaningful training compute is obtained by circumventing export controls; and inference demand is much smaller so more compute can go to training — though Moonshot has already paused new subscriptions because of K3, with the API still running. His mechanistic explanation: US labs spend a lot of energy on dramatic leaps that push the frontier, while Chinese labs focus more on catching up, and catching up is itself cheaper — just as a student model in distillation can surpass the teacher.
— Nathan LambertThe open ecosystem isn't holding, it's doubling down
The weekend after K3's release, Alibaba announced that the 2.4-trillion-parameter Qwen 3.8 would soon be released with open weights. The author stresses the weight of this signal: historically Alibaba kept its largest models exclusive to the cloud business's API, and this pivot shows Chinese companies aren't just maintaining the open-source status quo but doubling down on it — even if Qwen 3.8 trails K3 on benchmarks, the directional significance holds. He offers a verifiable inference: if Qwen 3.8 ships before the next Gemini generation, it pushes Google to eighth place on the smartest-model-lab ranking. DeepSeek V4 is also expected to graduate from preview. In his ranking of each player's peak model, DeepMind has only Gemini Flash 3.5 on the board; he describes seeing DeepMind and several US giants ranked so low as shocking, and thinks the X AI team deserves more credit.
— Nathan LambertBanning open source would create a defensive asymmetry
The author's position: if Claude Mythos were open-sourced today, the negative consequences would be relatively mild and the risk is overhyped; but that won't hold forever, so having the strongest models available in controlled, closed form a few months ahead of equally capable open models is a very safe equilibrium. The danger is regulation going off course. He cites Axios reporting: the Commerce Department considered adding several Chinese AI labs to the Entity List last year; the NSA and the White House Office of the National Cyber Director considered issuing threat advisories targeting Chinese labs; the White House considered an executive order requiring US companies to guarantee safety and assume liability for leaks before hosting Chinese models; and Commerce drafted rules targeting Chinese open models using supply-chain security authorities. The result would be a highly asymmetric state: America's best models carry guardrails on cybersecurity tasks, while global actors probe America's defences with excellent Chinese open models.
— Nathan LambertIn their own words · checked verbatim
The key fact is that either the open-to-closed or American-to-Chinese model performance gap has been reduced from the debated 6-9 months to something shorter, say 3-5 months.
Nathan Lambert0:00
Open-weight models are inherently decelerationist, and I’m continually surprised to see the so-called “accelerationists” so excited about open-weight models.
Dean Ball0:00
It is becoming clear that the Chinese labs are far more capital efficient. In a world where scaling laws dictate that intelligence is proportional to effective capital – which buys compute, data, & talent – that may be the greatest strength your AI industry could ever have.
Nathan Lambert9:23
Just as the student model can outperform the teacher in distillation generally (not limited to the adversarial distillation of the Chinese labs), an approach of “trying to catch up” rather than “invent the next paradigm” could lead to stronger models.
Nathan Lambert9:23
If we regulate open-weight models heavy-handedly, I suspect much of the world will be lulled into thinking we no longer need to act. All we would’ve done is slightly delayed the inevitable — open models will continue to cross all the key capability thresholds eventually and regardless of legality.
Nathan Lambert19:54
Figures
| K3 release date / weights release date | Released July 16, weights promised for July 27 | 0:00 |
| Open-versus-closed time gap | Narrowed from a contested 6-9 months to 3-5 months | 0:00 |
| K3 leaderboard positions | Vals AI #2, Artificial Analysis #3, Frontend Code Arena #1 | 0:00 |
| Compute available to a single OpenAI researcher (author's joking estimate) | Thousands of H100-equivalent machines | 0:00 |
| K3's MoE activation ratio | 16 of 896 experts activated | 9:23 |
| K3's scaling-efficiency gain over K2 | About 2.5× | 9:23 |
| Qwen 3.8 parameter count | 2.4 trillion parameters, to be released with open weights | 9:23 |
Glossary
- Kimi Delta Attention (KDA)
- An attention variant from the Kimi Linear paper that improves information flow along sequence length.
- Attention Residuals (AttnRes)
- An architectural change that improves information flow along model depth.
- Stable LatentMoE
- A mixture-of-experts framework used with extremely high sparsity (16 of 896).
- adversarial distillation
- Reverse-extracting capabilities from a closed frontier model's outputs to train your own model.
- decelerationist
- A force that compresses frontier labs' profits and funding, thereby slowing the pace of AI investment.
- Entity List
- A US Commerce Department control list; listed parties need a licence to obtain US technology and services.
How to listen
Investors watching the US-China model landscape, engineering leads choosing between open weights and closed APIs, and anyone doing AI export-control or open-source policy analysis.
Section 5, US regulatory developments — skip if you don't care about policy.