Werdmuller argues that as of mid-2026, every top open-weights model — Qwen, DeepSeek, Kimi, GLM — comes from Chinese labs, while American frontier labs (OpenAI, Anthropic, Google DeepMind) ship only behind APIs. He cites current leaderboards (LMSys, Artificial Analysis, coding evals) as evidence that the openable frontier is now entirely Chinese, and notes Meta's Llama has slowed with never-OSI-compliant licensing.
The editorial argues that closed-weights-as-safety was a coherent posture in an earlier era but no longer describes reality. Once a determined adversary can download Qwen or DeepSeek, API gatekeeping only constrains paying enterprise customers — not the actual threat model — making the safety justification a policy that binds compliant users while leaving bad actors unaffected.
The synthesis frames the current moment as a role reversal: in the 2010s, openness (TensorFlow, PyTorch, Kubernetes, React) was the American competitive advantage. In AI, that posture has flipped — Chinese labs now occupy the open-ecosystem position while American labs retreat behind APIs, a strategic loss whose implications the West has not fully absorbed.
The editorial observes that the 1,149-point HN thread was 'unusually free' of the typical 'benchmarks are gamed' hedging. Commenters accepted the ranking as fact and instead debated causes — signaling that within technical communities, Chinese open-weights supremacy has moved from contested claim to accepted baseline.
Ben Werdmuller's essay, which climbed to 1,149 points on Hacker News, makes a claim that a year ago would have read as trolling and today reads as bookkeeping: American AI is locked down and proprietary, and it's losing. The frontier of downloadable, modifiable, self-hostable model weights in mid-2026 is dominated by Chinese labs — Alibaba's Qwen, DeepSeek, Moonshot's Kimi, Zhipu's GLM, and a rotating cast of smaller outfits shipping specialized variants on a weekly cadence.
Meanwhile the American frontier — OpenAI, Anthropic, Google DeepMind — ships behind APIs. Meta's Llama line, once the standard-bearer for Western open weights, has visibly slowed, and its licensing was never OSI-compliant to begin with. The result is a market where the best models you can actually download, inspect, and run on your own hardware are, without exception, Chinese. That is not a rhetorical flourish. Pull up any current open-weights leaderboard — LMSys, Artificial Analysis, the various coding evals — and the top of the openable column is Qwen, DeepSeek, Kimi, GLM, in some order.
The HN thread underneath Werdmuller's post is unusually free of the usual "but benchmarks are gamed" hedging. The people arguing weren't disputing the ranking. They were arguing about *why*.
The conventional Western framing has been that closed weights are a safety posture: keep the model behind an API, monitor abuse, patch jailbreaks, avoid handing a weaponizable artifact to bad actors. It's a coherent argument. It also, in 2026, no longer describes reality. If a capable open-weight model exists anywhere in the world, safety-by-API-gatekeeping is a policy for your customers, not for the threat model. A determined adversary downloads Qwen or DeepSeek. Only your paying enterprise users are stuck on the guarded track.
The strategic inversion is worth naming plainly. In the 2010s, "open" was the American default — TensorFlow, PyTorch, Kubernetes, React, the entire CNCF landscape — and Chinese tech was caricatured as a walled fortress of Baidu clones. The AI era has flipped that script. Chinese labs are treating open weights as a distribution and standard-setting play, the same move Google made with Android against iOS a decade earlier. Every enterprise that fine-tunes on Qwen, every startup that builds on DeepSeek, every academic paper that benchmarks against GLM is one more knot tying the global developer ecosystem to a Chinese base layer.
What did the American labs get in exchange for their closure? Revenue, mostly. OpenAI's ARR is real. Anthropic's enterprise deals are real. But the moat those revenues were supposed to buy — capability lead — has been visibly compressing. GPT-5-class capability is now available in weights you can `wget`. The gap between the best closed model and the best open model, measured in months, has gone from "years" in 2023 to "a quarter, maybe two" in 2026, and the trendline is not in the closed side's favor.
There's a second-order effect the essay glances at but doesn't fully develop: the researcher pipeline. Open weights aren't just a product; they're a substrate for the next generation of papers. When every interpretability, alignment, and efficiency researcher on Earth is poking at Qwen internals because those are the internals they can see, the compounding advantage of *understanding your own model* accrues to the lab that released it. American frontier labs have effectively outsourced their own research community's substrate to Hangzhou.
If you're making architecture decisions in the second half of 2026, three things follow.
First: if your requirement is fine-tuning, on-prem, or air-gapped deployment, your shortlist is now almost entirely Chinese-origin models, and pretending otherwise is a procurement exercise, not an engineering one. Qwen3, DeepSeek-V3.x, and Kimi K2 are what your competent competitor is already running. The compliance conversation — "can we use a model with a Chinese-lab origin?" — is a real one in regulated industries and defense-adjacent work, but it's now a conversation you have to *have*, not one you can dodge by defaulting to Llama.
Second: the closed-API path is still the right call for a lot of workloads — anything where you want someone else on the hook for capacity, safety filters, and abuse response, and where the per-token economics work. But it's no longer the *only* serious path, and treating it as such is going to look, in retrospect, like sticking with Oracle in 2012 because "nobody gets fired for buying Oracle."
Third: watch the licensing fine print. Qwen and DeepSeek ship under permissive licenses that most legal teams can clear. Some of the smaller Chinese releases have more restrictive terms — non-commercial clauses, usage-reporting requirements, region carve-outs. "Open weights" is not a monolith, and the practical freedom varies more than the marketing suggests.
The interesting question isn't whether an American lab will release competitive open weights — Meta will keep trying, and there are credible rumors that at least one of the frontier labs is reconsidering. The interesting question is whether it will matter by the time they do. Standards get set early. Ecosystems get sticky. If the world's fine-tuning recipes, LoRA adapters, tokenizer conventions, and deployment tooling are all built around Qwen and DeepSeek architectures for another eighteen months, a late-arriving American open model doesn't reset the board — it just becomes one more option in a menu where the defaults have already been chosen.
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