Werdmuller argues American labs have deliberately chosen a closed, API-gated distribution model while Chinese labs like DeepSeek, Alibaba (Qwen), Moonshot, and Zhipu ship serious models under permissive open-weight licenses. He frames this not as China catching up but as America actively forfeiting the ecosystem battle through a strategic choice to keep frontier models behind rate limits and mutable terms of service.
The editorial points to Hugging Face trending and Ollama's registry being dominated by Qwen and DeepSeek variants as evidence that developers care about deployability, not benchmark supremacy. A 32B model that runs on a single H100 or quantizes onto a Mac Studio beats a 400B API-only model when it hits 90% of the quality on the workloads people actually run.
The synthesis characterizes Meta's Llama 4 reception as muted enough that its open-weights commitment now looks uncertain, and dismisses OpenAI's 2025 gpt-oss 20B/120B release as narrow and hedged behind a use-policy addendum rather than a genuine open commitment. This leaves the American stack proprietary by default with no credible open counterweight to the Chinese labs.
Ben Werdmuller's piece — which cleared 700 points on Hacker News and dominated the front page for most of a day — argues something that a year ago would have read as contrarian and now reads as obvious: the United States is losing the open AI race to China, and it's losing on purpose.
The scoreboard is easy to read. DeepSeek shipped V3 and R1 under an MIT-style license. Alibaba's Qwen team has released Qwen2.5, Qwen2.5-Coder, and Qwen3 variants across every size from 0.5B to 235B, all with commercial-use weights. Moonshot's Kimi K2 dropped as open weights. Zhipu's GLM-4.5 followed. Meanwhile the American frontier — GPT-5, Claude 4.5, Gemini 2.5 — is behind an API, a rate limit, and a terms-of-service document that reserves the right to change at any time.
Meta was the last major US lab shipping serious open weights, and Llama 4's reception was muted enough that the strategic direction now looks wobbly. OpenAI's gpt-oss release in 2025 was real but narrow — a 20B and 120B pair that ships behind a use-policy addendum, not a genuine open-weights commitment. The rest of the American stack is proprietary by default.
The usual framing of this — "China is catching up" — misses the actual dynamic. China isn't catching up. China picked a different distribution strategy, and the developer ecosystem is voting with `git clone`.
Look at Hugging Face's trending and download tabs on any given week this year. Qwen variants, DeepSeek variants, and their fine-tunes occupy most of the top slots. Ollama's model registry tells the same story: the models people actually pull to their laptops and their on-prem GPUs are overwhelmingly Chinese-origin. The reason is not ideology. It's that a 32B model you can run on a single H100 — or quantized on a Mac Studio — beats a 400B model you can only rent, if the 32B is 90% as good on your specific workload.
And the 32B is 90% as good on a lot of workloads. DeepSeek V3's reported training cost — famously around $6M for the final run — was met with skepticism, then partial confirmation, then a shrug. The exact number matters less than the direction: the cost of getting to *within striking distance* of frontier has collapsed. Qwen2.5-Coder-32B hits the top of coding leaderboards it has no business being on given its parameter count. GLM-4.5 does the same for agentic tool use.
The American labs' defense — "safety," "misuse risk," "national security" — has quietly stopped being convincing to the people actually shipping systems. Every capability that was supposed to stay locked behind an API is now downloadable from a Chinese lab within 90 days. The lockdown is not preventing proliferation. It is preventing American companies from being the ones downstream developers build on.
The compounding effect is what should worry US strategists. When Qwen becomes the default base model for a generation of fine-tunes, agent frameworks, and RAG stacks, the tooling ecosystem calcifies around it. vLLM optimizations, Unsloth training recipes, quantization schemes, MCP integrations — they get written against the open thing first. The closed thing becomes a specialty deployment target, not the substrate.
If you're currently building on Anthropic or OpenAI, the honest question is not "should I switch" — it's "what's my exit path if I have to?" A year ago the answer was "there isn't one, we're locked in." Today the answer is "Qwen2.5-72B on two H100s runs our workload at 60% of the quality for 15% of the cost, and Qwen3 closes most of that gap." That's not a migration you want to do, but it's a migration you *could* do — and having that option changes your negotiating posture with your current vendor.
For anyone building agent frameworks, coding assistants, or RAG pipelines: the default base-model assumption is shifting. The MCP server you write, the eval harness you build, the fine-tune recipe you commit — check whether it works against a Qwen or DeepSeek variant, not just against Claude. If it doesn't, you're building infrastructure on rented land.
For regulated industries — healthcare, finance, defense contractors, EU companies wary of US data residency — the open-weights option isn't a nice-to-have. It's the only way to run frontier-adjacent AI without shipping your customers' data to a third-party API. That was a theoretical constraint two years ago. It's a live procurement conversation now, and the answer that ends the conversation is "we self-host Qwen."
The unresolved question is whether any US lab reverses course. Meta could double down on Llama 5. A well-funded new entrant — a Mistral-style US play with actual open weights — is overdue. Absent that, the trajectory is set: the next generation of AI infrastructure gets built on Chinese base models by default, American labs become premium API vendors for the enterprises that can't or won't self-host, and "open" stops being a word American AI companies get to use with a straight face. The strategic self-own here is going to be studied for a 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