Alibaba Opens Qwen3.8-Max to the World — and Says the Weights Come Next Week

Alibaba made Qwen3.8-Max broadly available to developers worldwide today, through Alibaba Cloud's Model Studio APIs and through QwenWork, a workplace agent platform that entered public beta the same day. The company says the model's open weights follow next week.

The number in every headline will be 2.4 trillion parameters. That is the least interesting thing about this release. The interesting thing is the direction: as the South China Morning Post frames it, this is Alibaba's return to open-sourcing its top-tier models after keeping several recent flagships proprietary earlier this year.

What shipped today, and what didn't

Item	Available now	Still promised
API access	Model Studio, global	—
QwenWork agent platform	Public beta	—
Open weights	—	"Next week"
Licence terms	—	Unpublished

That bottom half matters more than it looks. "Open weights" is a phrase that spans everything from a true permissive licence to a release that forbids commercial use above a revenue threshold. Until the licence text exists, next week's release is a plan, not a commitment — and this is the second consecutive Qwen flagship to be announced well before it was obtainable. The model itself was first shown on 19 July, at the World AI Conference in Shanghai; today is when most people can actually call it.

The performance claim nobody outside Alibaba has checked

At the July announcement Alibaba positioned Qwen3.8-Max as second only to Claude Fable 5 — a claim carried by the trade press rather than one we have seen stated in a primary Alibaba document. No independent benchmark has been published, and the SCMP report announcing today's availability carries no scores at all.

Set that against the numbers on our own AI Model Leaderboard, drawn from the Artificial Analysis Intelligence Index as captured on 25 July:

Model	Index	Weights
Claude Fable 5	59.9	Closed
Kimi K3	57.1	Open, shipped 27 July
Qwen3.7 Max (predecessor)	46.0	Closed

For the claim to hold, Qwen3.8-Max would have to clear roughly 14 index points in a single generation and land above an already-open Kimi K3. Generational jumps of that size are rare enough that the burden sits squarely on the vendor. We have not added Qwen3.8-Max to the leaderboard, and will not until an independent score exists — a parameter count is not a capability measurement, and 2.4 trillion is in any case smaller than the 2.8 trillion of Kimi K3, whose weights have been downloadable since 27 July.

The awkward part of "second only to Fable 5"

There is a complication in benchmarking yourself against Claude specifically. In June we reported that Anthropic accused Alibaba of mass Claude distillation — training on outputs harvested from its models at scale. Anthropic has not, to our knowledge, produced account-level evidence publicly, and Alibaba has not conceded the allegation, so nothing here is settled.

But it does mean the comparison cuts in two directions at once. If a model is trained substantially on another model's outputs, scoring near that model is partly a description of the training data rather than an independent result. That is a reason to want third-party evaluation, not a reason to assume bad faith.

Why the reversal is the real story

Strip out the marketing and a pattern is visible. Over the past three weeks: Moonshot shipped Kimi K3 at 2.8 trillion parameters and then released the weights; DeepSeek pushed V4 Flash into the same price bracket as OpenAI's cheapest tier; and Alibaba is now walking a flagship back toward open release after a proprietary stretch.

That runs directly against the American picture we described when Jensen Huang's open-weights letter reached 50 signatories and Anthropic remained the last major holdout. The most notable US open-weight release of the period came from a startup — Mira Murati's Thinking Machines with Inkling — not from a frontier lab.

The counter-case is worth stating, because "China is winning on open weights" is too tidy. Releasing weights is cheapest for the lab that is second: it buys developer adoption and standard-setting influence that a leader would be giving away. Open weights are a competitive instrument here, not primarily an ideological commitment, and reading them as generosity misses what they are for. It also does not follow that an open model is a better one — that is precisely what the missing benchmark would tell us.

What to watch

Next week is the test, and it has three parts that can each be checked from outside. Whether the weights actually appear on schedule. What licence they carry, since a restrictive one would make "open weights" a marketing term. And whether an independent evaluator scores the model anywhere near the claim — at which point either the leaderboard changes or the claim quietly stops being repeated.