Developers can already try it through Alibaba's coding platforms; the evidence for the ranking has yet to be published.
Alibaba's Qwen team gave a first look at Qwen3.8-Max-Preview, a 2.4-trillion-parameter flagship the company publicly places behind only Fable 5, Anthropic's top model, and calls one of the most powerful systems available today, comparable to the field's leading frontier models.
Developers can reach the preview through Alibaba's coding tools, Qoder among them. Within the company's own line-up, it succeeds the previous Qwen flagship.
Nothing published so far allows an outsider to check the ranking. At the time of the preview there was no technical write-up, no model card, no benchmark table, no license file, no ordinary per-token price, and no date set for releasing weights.
The headline size leaves the practical question open. A sparse Mixture-of-Experts design underpins the model, and Alibaba has not said how many of the 2.4 trillion parameters are genuinely in use when it runs — the figure that bears on what such a model costs to serve.
The Qwen team says the model should beat its predecessor at writing code, building full-stack applications, analysing data and handling office work. Alibaba also plans an open-weight release soon, which would let developers download and modify it rather than depend on the hosted service alone.
One head-to-head run by Trilogy AI compared the preview with Kimi K3. Across 269 repository files in a coding-architecture benchmark, Kimi K3 scored 83 out of 100 after factual penalties against the new Qwen model's 80 — a three-point gap after blind review. Kimi finished faster on fewer tokens on the route tested. Qwen took the tool-use category 9 to 8, worked through fewer requests and tool calls, and had none of its 44 tool calls fail. Qwen drew a cleaner system boundary with a stronger replay and provenance record, the test found, while Kimi modelled the editing lifecycle, revisions and regeneration more completely.
The preview landed days after Moonshot AI released Kimi K3, which carries 2.8 trillion parameters, and while the World AI Conference was under way in Shanghai. Alibaba's Hong Kong-listed shares rose as much as 5.6% on July 20 before trimming to 5.15% at midday, against a 2.9% gain for the Hang Seng Tech Index. Zhipu's own run of losses in Hong Kong began earlier: close to 30% on July 17, then a further 14% on July 20.
In commentary carried by Morningstar, a research director at Gavekal Technologies said rivalry among large language model builders will probably sharpen until only a handful can still fund frontier work. On one developer forum the reaction split: many welcomed an open-weight contest among Chinese labs, while a smaller group doubted the Fable 5 placement and cast the model as tuned for benchmarks next to its competitors.
None of this warrants moving production workloads on a preview announcement. Five things have to exist first, and each is checkable when it does: a benchmark table, the active-parameter count, a licensed release with a real license file, published API pricing, and independent evaluation broader than the single matched run already reported. The open weights remain promised soon, with no date attached.