Werdmuller argues that America's proprietary API-only approach is a strategic mistake because it cedes the open ecosystem — where developers actually build, fine-tune, and study models — to Chinese labs. He points to Hugging Face dominance by DeepSeek, Qwen, Kimi, and GLM as evidence that closed weights don't protect national interests, they just guarantee that the substrate developers depend on will be Chinese rather than American.
Werdmuller frames the distributional argument: open weights enable on-device deployment for air-gapped, regulated, and cost-sensitive workloads that APIs categorically cannot serve. Beyond deployment, open weights let researchers fine-tune, distill, and publish on the models — every derivative work and arxiv paper compounds into a knowledge base that entrenches whichever ecosystem hosts the weights.
Werdmuller contends that Washington's export-control regime and the frontier labs' closed-weights lobbying rest on a flawed premise — that keeping weights secret protects American AI leadership. He argues the opposite is true: DeepSeek-R1's release forced a repricing across the reasoning-model market and its architecture was studied inside every US lab within 48 hours, showing that secrecy delays nothing and only hands the open ecosystem to competitors.
Werdmuller identifies Llama as the lone American open-weights entrant at frontier scale and notes it has visibly stalled — Llama 4 shipped to muted reception and Meta reportedly restructured its GenAI org. Without a functioning US open-weights lab, the community's default fine-tuning base becomes Qwen or DeepSeek by process of elimination, not by ideological preference.
Ben Werdmuller's essay, which hit #1 on Hacker News with over 1,000 points, crystallized a shift that's been building for eighteen months: the world's most capable openly-licensed AI models are now overwhelmingly Chinese. DeepSeek-V3 and R1, Alibaba's Qwen family, Moonshot's Kimi K2, and Zhipu's GLM series ship weights on Hugging Face under permissive licenses. Meanwhile OpenAI, Anthropic, and Google — the American frontier — keep their crown-jewel models behind rate-limited APIs, with terms of service that forbid using outputs to train competitors.
The scoreboard on Hugging Face is unambiguous: Chinese labs occupy most of the top slots for downloads, fine-tunes, and derivative works. Qwen alone has spawned tens of thousands of community variants. DeepSeek-R1 forced a repricing across the entire reasoning-model market when it landed in January, and its architecture papers were read like samizdat inside every US lab within 48 hours. Meta's Llama, the one American open-weights holdout at frontier scale, has visibly slowed — Llama 4 shipped to muted reception and Meta reportedly restructured its GenAI org.
The US policy stance points the opposite way. Export controls target Chinese access to Nvidia chips. The reasoning inside Washington — and inside the frontier labs that lobby it — is that keeping weights closed is a national-security asset. Werdmuller's argument is that this reasoning has it backwards: closed weights don't protect anything, they just guarantee that the open ecosystem developers actually build on will be someone else's.
The framing that matters here isn't ideological, it's distributional. When weights are open, three things happen that don't happen with an API. Developers can run the model on their own hardware, which means air-gapped, regulated, or cost-sensitive workloads become tractable. Researchers can fine-tune, distill, and study the model, which compounds into a knowledge base — every arxiv paper written against Qwen makes Qwen more useful. And downstream products get built on the assumption the weights will still exist in five years, which no API can promise.
Every serious open-source AI project of 2025 — vLLM optimizations, quantization work, agent frameworks, on-device inference — is being developed and benchmarked against Chinese base models first, American models second if at all. That's the leading indicator. The lagging indicator will be enterprise adoption, and it's already visible: banks, defense contractors, and healthcare shops that can't send data to OpenAI are quietly standing up Qwen and DeepSeek clusters. Some of them are the same institutions Washington claims to be protecting.
The cost story reinforces this. DeepSeek-V3 was trained for a reported $5.6M in compute. Whether or not that number captures the full picture (it almost certainly doesn't include the R&D salaries or the failed runs), it set a public anchor that reframed the trillion-dollar CapEx narrative coming out of Menlo Park. When a model that matches GPT-4o class performance can be trained for the price of a Manhattan brownstone and then given away, the moat argument for API-only distribution gets thinner.
The counter-argument from the closed-weights camp is safety: uncontrolled weights can be fine-tuned to remove guardrails, and once they're out, they're out. This is technically true and strategically irrelevant. The safety argument only works if you're the only lab in the world capable of building the model — the moment a comparable open model exists, your closed release provides zero marginal safety. That threshold was crossed sometime in 2024. Refusing to release weights in 2026 doesn't prevent misuse; it just concedes the developer ecosystem.
There's also a quieter dynamic inside the labs themselves. Every talented researcher who joins Anthropic or OpenAI now knows their best work will live behind an API their friends can't run. That's a recruiting headwind that didn't exist two years ago. Meanwhile a Qwen or DeepSeek paper drops and the author's name is attached to something the entire community will use for a decade.
If you're picking a model to build on in mid-2026, the calculus has genuinely changed. For most application workloads — RAG, structured extraction, agentic tool use, code assist — a well-tuned Qwen or DeepSeek deployment is now within a few percentage points of the closed frontier at a fraction of the marginal cost. The gap that remains is in the very top end of reasoning and in the polish of the closed models' tool-use reliability, and that gap is closing quarter over quarter.
The practical move for senior engineers isn't to rip out Claude or GPT — it's to make sure at least one production path in your stack runs on open weights, so you have real optionality when pricing, policy, or geopolitics shift. That could be a self-hosted Qwen for the batch pipeline while the interactive tier stays on Anthropic. It could be a DeepSeek distillation for the on-device feature. What you don't want is a stack where every AI call terminates at a single vendor's rate limiter with no fallback that isn't a total rewrite.
For teams building AI products for regulated customers — health, finance, government, EU data-residency shops — the open-weights option isn't optional anymore, it's the only option that closes procurement cycles. That's a bigger market than most SF founders model, and it's structurally biased toward whoever ships weights.
The most likely 2027 scenario is a bifurcated market: American closed models keep the consumer chat and enterprise white-glove tiers, Chinese open weights take the developer platform layer, and the open ecosystem — tooling, fine-tunes, research — increasingly speaks Chinese by default. That's not a prediction about geopolitics, it's a prediction about where the next generation of AI engineers will do their first fine-tune. Whichever lab owns that moment owns the decade.
I’m suspicious of some quotes here, “80% of startups using Chinese models,” doesn’t seem quite right to me. I just interviewed at several startups and they were all using the US models. Maybe they have some minor use of Chinese models but the bread-and-butter of most of these businesses model use is
This is a very strange article considering that Llama, the mother of all open-weight models, has led to anything but success for Meta.Also, enterprises don't give a rip if models are open. They care about zero data retention (and sticking with whatever vendor they're already using).This bl
I do think open-weights models are going to "win" in the sense that they're probably going to be dominant when the hardware to run them becomes affordable. (which might be a while). Although I guess you could probably rent the GPU's yourself to hypothetically save on costs. (I&#x
I do not understand the logic going into these companies. Flagrantly violate all IP in Human history, essentially claiming domain over the heritage of Humanity... And... Try to privatize it? When the technology -- and data -- are both public domain to begin with?It is ming-boggling stupidity. If the
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
The lesson of the last 50 years of the computer and software marketplace is that free and low-end eventually wins.- PCs destroyed minicomputers. Mainframes survive, but serving a much tinier portion of the market than they used to.- PC office productivity software destroyed expensive professional pr