Alibaba announces the biggest open AI model release of the year so far. Press coverage, superlatives, "open source" in every headline. A few days later, developers look inside the downloaded files — and notice the thing can't actually do everything the announcement promised.
Welcome to Qwen3.8-Max, Alibaba's new flagship. It really is huge. Whether it's "open" is... complicated.
What Qwen3.8-Max is supposed to do
A few numbers to place it:
- 2.4 trillion parameters. The dials tuned during training — roughly: the more of them, the more knowledge fits in. That makes Qwen3.8-Max one of the largest AI models around, well ahead of most US competitors.
- 95 billion active parameters per query. That comes down to the architecture: a so-called Mixture-of-Experts model only uses a fraction of its knowledge for any given request — faster and cheaper than a model that always fires up everything at once.
- 1 million tokens of context. A token is roughly a syllable — so the model can hold entire books in mind at once. At least the cloud version can.
And that's exactly where the catch starts.
The catch: open doesn't mean open
Alibaba promised free model weights (the trained numerical values you can run yourself) for the week of August 10 — and actually delivered. A download has been sitting on Hugging Face since then, free, no credit card required.
Except: read the fine print, and you find a noticeably stripped-down version. No image input, no video, none of the 1-million-token context length. The full capabilities — image understanding, the massive context window — currently exist only behind Alibaba's paid cloud API. On top of that comes a custom-built "Qwen3.8-Max License" instead of one of the usual open licenses. It's like a car dealer offering a "free" car — just without the engine, that's sold separately.
To be fair: downloading a 2.4-trillion-parameter text model for nothing is still remarkable. Just not quite what the headlines promised.
Who can actually run this thing?
Almost nobody, honestly. 2.4 trillion parameters don't fit on a home PC, and most company servers can't handle it either. Running it yourself takes data-center hardware in the six-figure range — nothing like "run it on your own PC," the way you could with smaller open models.
What's free, and what costs money?
To try it without downloading anything, Qwen Chat is enough — Alibaba's free web chat, works like ChatGPT or Claude in the browser, no installation needed.
If you want in through the programming interface (API), signing up with Alibaba Cloud currently gets you a free quota of roughly one million tokens. After that it currently costs around 2 US dollars per million input tokens and 6 dollars per million output tokens — cheaper than most US flagship models, but not free.
Is this for me?
If you just want an AI to chat with: Qwen Chat is enough, free, no technical knowledge required. For anything beyond that — self-hosting, API integration, the full feature set with images and long context — you're firmly in developer territory, and you'll end up paying for the cloud version anyway.
My take
Qwen3.8-Max is technically impressive — 2.4 trillion parameters is nothing to sneeze at. But "open" is being interpreted generously here: the free version is the trimmed-down one, and the genuinely interesting capabilities cost extra. Before you share the next "revolutionary open AI" headline, take a moment to check what's actually in the download. Usually worth the second look.