Qwen3.8-Max: 2.4T parameters, and no benchmarks to check it

Alibaba previewed Qwen3.8-Max on July 19 at the World AI Conference in Shanghai: a 2.4-trillion-parameter sparse mixture-of-experts model handling text, images, video, and documents. The Qwen team called it second only to Anthropic's Fable 5. What they did not ship: a model card, a license, a benchmark table, or the active-parameter count. If you're deciding whether Qwen3.8-Max belongs in your stack, the interesting story is the missing evidence, not the parameter count.

What actually happened

The preview landed as two posts on X plus a working paid endpoint — no formal blog post, per MarkTechPost's writeup. Alibaba says the model will become open-weight but hasn't announced a date or a license. The claim is that it outperforms Qwen3.7-Max on coding, development, data analysis, and office tasks. Markets liked it: Bloomberg reported Alibaba shares rose as much as 5.4% on the news.

Access runs through Alibaba's Token Plan, a credit-based subscription with individual tiers from 39 yuan/month up to 499 yuan/month, alongside team seats. Credit consumption is running at a 90% launch discount — meaning preview economics are promotional economics, not the number you'll budget against.

Two things to hold onto. The active-parameter count is the number that determines serving cost on a sparse MoE model, and it's undisclosed. And a 2.4T model at 4-bit precision needs roughly 1.2TB just for weights, which puts "self-host it" well outside a single GPU and probably outside your rack.

Why an unverified model release matters for your business

We keep the model layer swappable precisely so that a launch like this is a config change, not a project. But swappable doesn't mean you swap on a press release.

The discipline is boring and it works: run the model against your own workload before it goes near production. Not MMLU. Your ten hardest real tickets, your actual product-description generation, your actual invoice extraction — scored the way you'd score a contractor. Vendor benchmarks are marketing even when they exist. Here they don't exist at all, so a claim of "second only to Fable 5" is a claim, full stop.

Second: price against list, not promo. A 90% launch discount is a customer-acquisition cost, and it expires. If the business case for switching only works at 10% of list price, you don't have a business case — you have a trial.

Third: "will become open-weight" is not open-weight. No license means no portability guarantee. Treat it as a hosted API until the weights and the license actually land.

Sources: MarkTechPost, Bloomberg.