Alibaba previews Qwen 3.8-Max, a 2.4-trillion-parameter model it ranks second only to Fable 5
Alibaba's Qwen team used WAIC in Shanghai to preview Qwen 3.8-Max, a 2.4 trillion-parameter multimodal model it says trails only Anthropic's Fable 5. The claim arrives without benchmarks, a preview build, and an open-weight release promised only "soon."
Alibaba used the World Artificial Intelligence Conference in Shanghai to preview Qwen 3.8-Max, a 2.4 trillion-parameter model the Qwen team describes as “second only to Fable 5.” It is the first Qwen model above a trillion parameters to handle images, video, and documents alongside text, and the loudest claim in the announcement is also the least substantiated: Alibaba published no benchmarks, no model card, and no count of how many parameters are active at inference.
What Alibaba previewed
The July 19 post positions Qwen 3.8-Max as a broad step up from Qwen 3.7-Max, “especially in coding and complex productivity tasks such as full-stack development, data analysis, and office workflows.” For now it exists as a preview build reachable through Alibaba’s Token Plan subscription and its Qoder and QoderWork developer surfaces, priced at 10 percent of the standard rate during the trial window. Open weights are promised “soon,” which would break Alibaba’s habit of keeping its Max-tier models closed while shipping the smaller Qwen line under open licenses.
Two numbers frame the model. At 2.4 trillion parameters it is the second-largest publicly disclosed model behind Moonshot AI’s Kimi K3, a 2.8 trillion-parameter release that landed three days earlier, and it is far larger than Qwen 3.7-Max, which scored 56.6 on the Artificial Analysis Intelligence Index in May and sat well below Claude’s flagship on public leaderboards. A jump of that size normally signals a real capability change. Without an activated-parameter figure, though, there is no way yet to tell whether Qwen 3.8-Max is a dense model or a sparse mixture-of-experts design where only a fraction of those parameters fire per token, and that distinction drives both cost and speed.
Where this lands in the market
The timing reads as a direct answer to Kimi K3. Moonshot has been converting interest in K3 into paying users by gating access through its own app and API even while promising open weights, and it is reportedly steering toward an IPO. An open-weight Qwen model in the same size class would undercut that funnel by handing teams the artifact rather than metered access. For the broader open-model camp, a 2.4 trillion-parameter release from a vendor with Alibaba’s distribution is another sign that the frontier and the open tier are converging faster than they were a year ago.
For a working developer, the practical takeaway is narrower than the headline. A 2.4 trillion-parameter model is not something most teams will self-host regardless of license, so the near-term question is what the hosted preview actually does on real coding and agent workloads at a tenth of standard pricing. The “second only to Fable 5” line is a marketing frame until an independent evaluation puts a number behind it, and Alibaba’s own history of strong open small models does not transfer automatically to a new flagship.
What’s worth watching
- Benchmarks from someone other than Alibaba. The ranking claim is unverified. Whether Artificial Analysis, LMArena, or a coding index reproduces a top-two placement is the signal that decides if this is a Fable 5 competitor or a press-release number.
- The open-weight release and its license. “Soon” with no date and no license terms is the load-bearing promise here. A permissive open-weight drop at this scale would matter more to the ecosystem than the preview does; a delayed or restricted one would make this a closed Max model like the others.
- Active-parameter disclosure. Until Alibaba says how much of the model runs per token, the cost and latency story is unknown, and that is what determines whether teams can afford to leave it in an agent loop.
The honest read is that Alibaba announced ambition on schedule for a conference stage and left the verification for later. The preview is real and worth testing against your own workload, but the ranking is a claim, not a result. Stackmaven will revisit Qwen 3.8-Max once independent benchmarks and the open-weight terms are on record, on or around October 18.