Scroll the Qwen account this morning and there's the flex: Qwen3.8, 2.4 trillion parameters, going open-weight soon, plus a line calling it second only to Fable 5. So is it out? Sort of. What actually shipped on July 19 is Qwen3.8-Max-Preview, a hosted preview you can poke at through Alibaba's Token Plan and its Qoder tools. The weights aren't on Hugging Face. That 2.4T figure, the frontier-rival claim, none of it ships with a model card or a benchmark table. It's one tweet from the Qwen team and a preview endpoint. We went to split what you can use this afternoon from what's still just a promise.
The short answer
Alibaba announced Qwen3.8 on July 19 and put a hosted Qwen3.8-Max-Preview up on its Token Plan and Qoder tools. You can prompt it now. The 2.4 trillion parameter count, the open weights and the claim that it trails only Fable 5 are all promises attached to a single tweet. No model card, no benchmarks, no date. Here’s the line between what shipped and what didn’t.
What actually went live
A preview you can send prompts to. That’s the solid part. On July 19 the Qwen team posted that Qwen3.8 is “launching and going open-weight soon”, and in the same breath pointed people at Qwen3.8-Max-Preview on Alibaba’s Token Plan, Qoder and QoderWork. So if you have access to those, the model is reachable right now.
Note the name. Max-Preview, not the open release. That’s the same two-step Qwen has run before: a hosted “Max” preview goes up first, the downloadable weights follow later, sometimes weeks later. What’s in your hands today is the closed, hosted version. Useful, but it’s not the thing the headline is really selling.
The one word doing a lot of work in that tweet is “soon”. No date. And “soon” from a model lab has a wide range.
The numbers Qwen is quoting itself
Here’s where we slow down. The tweet carries two big claims, and both read like spec-sheet facts. They aren’t. Not yet.
First, the size: 2.4 trillion parameters, described as “massive” and “continuously evolving”. That figure comes from the launch post, full stop. There’s no model card, no technical report, and crucially no active-parameter count. A 2.4T total on a mixture-of-experts model might light up only a fraction per token, which is what actually decides your latency and cost. Qwen hasn’t said. For scale, Qwen3.6-Max-Preview was reported around 1T total, so 2.4T is roughly a 2x jump on paper.
Second, the ranking. Qwen says the model is “compatible to leading frontier AI models, second only to Fable 5”. That’s a strong claim, and it’s marketing, not a benchmark you can pull up and rerun. Nobody outside Alibaba has scored it. A launch-day boast from the lab that built the model is the least independent number in the room. It might hold. We’d wait for someone neutral to run Fable 5 against it before repeating the line.
What “continuously evolving” costs you
That phrase in the tweet is easy to skim past, and it matters. A preview that’s “continuously evolving” is a moving target. The weights behind the endpoint can change between Tuesday and Friday, which means a prompt you tuned this week can drift next week. Fine for kicking the tyres. Bad for anything you’re putting into production or benchmarking for a report.
So treat Qwen3.8-Max-Preview as exactly what it says on the tin: a preview. Throw your real prompts at it, get a feel, but don’t wire it into a pipeline and don’t publish “Qwen3.8 scores X” off a preview that’s admittedly still changing.
The open-weight part, and the hardware wall
Here’s the honest read on “open-weight soon”. When the weights do land, a 2.4T model is not something you run on a desk. It’s the same wall we keep hitting: this month Thinking Machines Inkling shipped as “open”, and open meant a rack of big GPUs, not your laptop. A 2.4T total will be heavier still. “Open” here means “you can self-host if you own a small cluster”, not “you can pull it down tonight”.
If what you actually want is Qwen weights that run on hardware you own, that’s a different model. The smaller Qwen 3.7 offline build is still the sensible pick, and if you’re weighing it against the field, we put it up against GLM in Qwen 3.7 Max vs GLM 5.2. Qwen3.8’s open release, whenever it arrives, is aimed at the datacenter, not the home lab.
The honest read
I’m curious about Qwen3.8, and I’m not going to repeat its scoreboard as fact. Both hold at once.
What’s real today is small but usable: there’s a Qwen3.8-Max-Preview you can prompt through Alibaba’s tools, and that’s it for confirmed. The 2.4T count, the “second only to Fable 5” line, the open weights, the date, all of it is running ahead of anything Qwen has actually published. So do the boring thing. If you have preview access, feed it the prompts that matter to your work and see how it holds up. Then wait for the weights and a real model card before you trust a single decimal point. The tweet is Qwen’s marketing. The prompts on your desk are the only test that pays your bills.
Sources: Qwen (@Alibaba_Qwen) on X, July 19, 2026, the launch post announcing Qwen3.8, the 2.4T figure, the open-weight plan and the “second only to Fable 5” claim. Availability of Qwen3.8-Max-Preview on Alibaba Token Plan, Qoder and QoderWork is per the same announcement. Qwen3.6-Max-Preview’s reported ~1T total parameters and Qwen3.7-Max list pricing are per Artificial Analysis and OpenRouter. The 2.4T parameter count, the benchmark ranking and any Qwen3.8 pricing are as claimed by Qwen, are not from an official model card, and are not independently reproduced.
Frequently asked questions
Is Qwen3.8 released?
Partly. On July 19, 2026, Alibaba announced Qwen3.8 and made a hosted Qwen3.8-Max-Preview available through its Token Plan, Qoder and QoderWork surfaces, so you can send it prompts today. The open weights are not out. Qwen says they are coming soon, with no date and nothing on Hugging Face yet.
How big is Qwen3.8?
Qwen says 2.4 trillion total parameters. That number comes from the launch tweet, not a model card, and Qwen has not disclosed how many parameters are active per token. For scale, Qwen3.6-Max-Preview was reported at roughly 1T total, so 2.4T would be a bit over a 2x jump. Treat it as a claim until the weights and a technical report land.
Can I download the Qwen3.8 weights?
Not yet. Qwen calls Qwen3.8 open-weight and says it is going open soon, but as of July 19 the only access is the hosted Qwen3.8-Max-Preview via Alibaba. If you want Qwen weights you can run on your own hardware today, the smaller Qwen 3.7 releases are still the honest answer.
Is Qwen3.8 better than Claude Fable 5?
Qwen claims it is second only to Fable 5 among current models. That is Qwen own wording, not a published benchmark table, and nobody outside Alibaba has reproduced it. So we would not repeat it as fact. Run your own prompts against the preview and judge from there.
How much does Qwen3.8-Max-Preview cost?
Alibaba had not posted per-token rates for Qwen3.8-Max-Preview as of July 19. For a rough anchor, Qwen3.7-Max lists around $2.50 in and $7.50 out per million tokens, and Qwen3.6-Max-Preview came in cheaper. Do not budget on those for 3.8 until Qwen publishes official numbers.