Artificial Intelligence

Alibaba Withheld the Number That Sets Your Cost

Alibaba previewed Qwen3.8-Max at 2.4 trillion parameters and claimed second place behind Fable 5, without publishing benchmarks, an active parameter count or an open-weight date. What a preview priced at 90 percent off is actually telling you.

Artificial IntelligenceBy Servola Tech Desk2026-07-204 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Alibaba Withheld the Number That Sets Your Cost

Key takeaways

Alibaba previewed Qwen3.8-Max on 19 July at the World AI Conference in Shanghai with 2.4 trillion total parameters, its first Qwen multimodal model above one trillion, and it did not disclose how many of those parameters are active at inference, which is the figure that actually determines what a query costs you.
Treat a total parameter count as marketing until the active count is published. In a sparse mixture-of-experts model only a fraction of parameters run per query, so 2.4 trillion total is compatible with almost any real cost, and Alibaba has published the impressive number and withheld the useful one.
Alibaba says the model is second only to Claude Fable 5, and released no benchmarks to support it. A vendor ranking itself against a named competitor without published evaluations is a claim, not a result, and the independent leaderboards have not scored it yet.
Do not move a roadmap onto this model yet. The preview runs through Alibaba's Token Plan, Qoder and QoderWork at 10 percent of standard price, and a 90 percent discount on an unbenchmarked preview is a request for evaluation traffic, not a production offer.
The open-weight promise has no date attached. Alibaba said weights are coming soon without committing to when, which matters because the comparison everyone is drawing is against Moonshot's Kimi K3, whose weights have a stated date of 27 July.

What Alibaba announced in Shanghai

On 19 July, at the World AI Conference in Shanghai, Alibaba's Qwen team previewed Qwen3.8-Max, the next flagship in the Qwen family. The headline specification is 2.4 trillion parameters, and it is the team's first multimodal model above one trillion, handling text, images, video and documents in one system.

Alibaba positions it as second only to Claude Fable 5 and says it improves on Qwen3.7-Max in coding, full-stack development, data analysis and office workflows. A preview is available through Alibaba's Token Plan, Qoder and QoderWork at 10 percent of the standard price, and the company has promised an open-weight release soon without naming a date.

The missing figure is the one that prices your workload

Why it matters: 2.4 trillion is the number that gets reported and the number that tells you least. Large frontier models of this shape are sparse mixture-of-experts systems, which means only a subset of the parameters is engaged for any given query. The total describes how much model exists. The active count describes how much of it runs, and therefore how much compute each request consumes.

Alibaba has not disclosed the active count. That single omission makes the announced size unusable for planning: a 2.4 trillion parameter model with a small active fraction can be cheaper to serve than a much smaller dense model, and one with a large active fraction can be far more expensive. Anyone sizing a budget from the headline figure is working from a number that was chosen to impress rather than to inform.

A ranking claim with nothing behind it yet

The second-place claim deserves the same scrutiny. Alibaba has named a specific competitor, Claude Fable 5, and placed itself directly beneath it, while publishing no benchmark results to support the placement.

That is the reverse of the pattern worth trusting. When Moonshot released Kimi K3, its own announcement placed the model second overall, and independent testing on Arena.ai then ranked it first for frontend code. A vendor whose own claim is more modest than the independent result has given you a usable signal. A vendor whose claim arrives without any independent result has given you a press line. Until the public leaderboards score Qwen3.8-Max, the ranking is Alibaba's opinion of Alibaba.

Read the discount as what it is

The bottom line: a preview at 10 percent of list price is an invitation to test, not an offer to build on. The pricing is doing a job, which is to get real workloads onto an unproven model quickly enough to generate the evaluation evidence the launch did not include.

There is nothing improper about that, and for a European team with spare evaluation capacity it can be a cheap way to learn something. The mistake is treating the discounted rate as the economics of the model. Preview pricing ends, the standard rate is ten times higher, and any business case built on the promotional number inverts the moment the preview closes. If you test it, test it on your own workload and record what it costs at full list, not at the introductory rate.

Soon is not a date, and the comparison has one

The open-weight promise is the part with the clearest consequence for anyone weighing independence from a hosted API. Alibaba says weights are coming soon and has not committed to when. Moonshot has said Kimi K3's full weights arrive on 27 July.

For an operator who cares about being able to run a model on infrastructure it controls, those two statements are not comparable. One is a plan with a date you can put in a calendar and hold a vendor to. The other is an intention. If your reason for watching Chinese frontier models at all is that open weights let you self-host, inspect and leave, then a release without a date does not yet advance that goal, however large the parameter count attached to it.

Frequently asked questions

How does Qwen3.8-Max compare with Kimi K3?

On published evidence, it does not yet compare. Kimi K3 has independent leaderboard results and a stated open-weight date of 27 July. Qwen3.8-Max has a vendor claim of second place behind Fable 5, no published benchmarks, no disclosed active parameter count and no open-weight date.

Why does the active parameter count matter more than the total?

Because in a sparse mixture-of-experts model only part of the network runs for each query. The total tells you how large the model is, while the active count tells you how much compute a request consumes and therefore what it costs to serve. Without the active figure, the total cannot be turned into a cost estimate.

Should we test the preview at the discounted price?

Testing is reasonable if you have evaluation capacity to spare, but record results against the full list price rather than the 10 percent preview rate. Preview pricing is temporary, and a business case built on it will not survive the preview ending.

A launch that leads with total parameters, skips benchmarks, omits the active count and dates the open weights as soon has told you what it is comfortable being measured on. Wait for the independent scores and the release date, then decide.

AlibabaQwenOpen weightsMixture of expertsKimi K3WAICModel benchmarksInference cost

Nvidia's Real Lock-In Just Went Open Source

Alibaba's T-Head opened its SAIL software stack to challenge Nvidia's CUDA lock-in. For a European operator the useful lever is not the chip but your own framework layer.

4 min read

AI Policy

One Phone, a Different AI Behind Each Border

China cleared Apple Intelligence on 15 July, but in China it runs on Alibaba's Qwen, not Apple's own model. The real lesson for owners is model routing by jurisdiction.

3 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.