This week, a Reddit post titled “New Qwen models coming” shot to the top of Hacker News — 608 points, 445 comments — all for a model that has not been announced, has no release date, and exists only as a secondhand anecdote from a conference in Paris. An Alibaba representative, the post claimed, said Qwen 3.8 would drop in August, with Qwen 4 to follow in September. The thread was not skeptical. It was excited.
This is not a story about Chinese AI catching up to American labs. It is a story about how the open-source AI movement has been quietly rewired into a permanent anticipation machine — and how willingly the developer community has played along.
The Model That Was Always Next
Review the recent timeline. Qwen 3.6 launched in April 2026. The 3.6-Max and 3.6-Plus variants were kept proprietary — paid API access only, no open weights. In May, at the Apsara Summit in Hangzhou, Alibaba announced Qwen 3.7-Max as a “preview.” Two variants went live on chat.qwen.ai. OpenRouter listed them at $2.50 per million input tokens and $7.50 per million output tokens. As of May 23, no open weights had appeared on Hugging Face.
Now, before the 3.7 weights have materialized — before anyone outside Alibaba can run the model on their own hardware — the conversation has already moved to 3.8. And 4. The community that once demanded open weights now settles for previews. The previews are free to try, cheap to call via API, and just good enough to keep you from looking elsewhere. The weights, meanwhile, remain a promise.
One developer who attended the Paris conference told me in a direct message afterward: “He mentioned 3.8 and the room lit up. Nobody asked when 3.7 weights were dropping. I don’t think anyone cares anymore.” That is not cynicism. It is market research.
The Funnel Is the Product
There is a term for this in consumer software: the perpetual beta. Gmail wore that label for five years. But Gmail was free. Alibaba’s model services, by contrast, are on track to generate 30 billion yuan in annualized recurring revenue, according to the company’s own disclosures. The previews are not charity. They are lead generation.
The pattern is by now familiar. Release a preview. Let the benchmarks circulate. Let the community build tutorials, fine-tunes, and breathless comparison threads. By the time anyone asks about open weights, announce the next version. The cycle resets. The community, trained to chase the new, forgets to ask about the old.
This is not a criticism of Alibaba specifically. The entire frontier-model industry has converged on the same playbook. What makes the Qwen story instructive is the speed. Four major version increments in roughly four months — 3.5, 3.6, 3.7, 3.8 — is not a research cadence. It is a marketing cadence. The version numbers are doing the work that press releases used to do.
What Gets Lost When Everything Is a Preview
The practical consequence is that “open-source AI” is becoming a spectrum rather than a category. At one end, you have truly open models — weights downloadable, licenses permissive, no API required. At the other, you have what the industry now calls “open-weight-later” or “preview-first” releases. The distinction matters because the downstream ecosystem — the startups, the researchers, the hobbyists running models on consumer hardware — cannot build on a promise.
When Meta released Llama 3 in April 2024, the weights were available immediately. You could download them, quantize them, and run them on a MacBook the same day. That immediacy created an ecosystem. The preview model, by contrast, creates dependency. You integrate the API. You build your product on someone else’s rate limit. You wait.
The Hacker News thread about Qwen 3.8 is, in this light, a kind of collective suspension of disbelief. Six hundred people upvoted a rumor about a model that may or may not ship with open weights, from a company whose last two “Max” variants never did. The enthusiasm is genuine. The object of the enthusiasm is imaginary.
The Anticipation Economy
None of this is illegal, or even particularly deceptive. Alibaba has not promised open weights for 3.7 or 3.8. The community has simply assumed that “Qwen” means “open,” because it once did. The Qwen 2.5 series shipped weights. The brand equity from those releases is still paying dividends, even as the actual release strategy has shifted.
The broader lesson is that anticipation is now a product in its own right. The HN thread, the Reddit rumors, the conference whispers — they generate engagement, mindshare, and API signups without requiring a single weight to be committed to a public repository. The model you can actually run is always one version behind the model you are excited about. That gap is where the money is.
Alibaba is spending heavily on AI, and the spending is working. The preview models rank among the top systems on LM Arena. The technology is real. But the community that celebrates every new version number might pause to ask: what exactly are we celebrating? A model we can use, or a model we can anticipate? The difference is larger than a decimal point.