Alibaba’s Qwen3.8-Max: 2.4 Trillion Parameters, No Benchmarks Yet
Alibaba previewed Qwen3.8-Max at WAIC 2026 with 2.4 trillion parameters but zero benchmarks. Here's what we know and what to wait for.
Alibaba previewed Qwen3.8-Max-Preview on July 19, 2026, at the World Artificial Intelligence Conference in Shanghai. The model carries 2.4 trillion parameters, making it the first Qwen model above 1 trillion parameters that can process images, video, and documents alongside text. Alibaba's Qwen team posted on X that the model is "second only to Fable 5," referring to Anthropic's Claude Fable 5. No benchmark scores, model card, or independent evaluation have been published to support that claim.
What happened
| Detail | Value |
|---|---|
| Announced | July 19, 2026, at WAIC in Shanghai |
| Model name | Qwen3.8-Max-Preview |
| Parameters | 2.4 trillion |
| Modalities | Text, images, video, documents |
| Claimed ranking | Second only to Claude Fable 5 (Anthropic) |
| Benchmarks published | None |
| Trial pricing | 10% of standard rate via Token Plan, Qoder, and QoderWork |
| Open weights | Promised “soon,” no date or license terms set |
Alibaba released the model as a preview build, not a finished product. The company described it as “continuously evolving” and has not published a model card, an activated-parameter count, or any task-level comparison against its predecessor, Qwen3.7-Max.
The timing is pointed. Chinese startup Beijing Moonshot AI released Kimi K3, a 2.8-trillion-parameter model, just three days before this announcement. Both releases push Chinese AI into a parameter scale that, until recently, only the largest US labs operated at. Our earlier coverage of Kimi K3’s market impact gives useful context on what that rivalry looks like in practice.
How does Qwen3.8 compare to its predecessor?
| Model | Parameters | AI Analysis Intelligence Index | LMArena position | Benchmarks at launch |
|---|---|---|---|---|
| Qwen3.7-Max (May 2026) | Not published | 56.6 | Well below No. 1 | Full set published |
| Qwen3.8-Max-Preview (July 2026) | 2.4 trillion | Not published | Not yet scored | None |
| Claude Fable 5 (Anthropic) | Not stated in source | Not stated in source | No. 1 | N/A |
The contrast between the two Qwen launches is striking. Qwen3.7-Max shipped in May with a full published results set. Qwen3.8-Max-Preview arrived with none. For a claim as strong as “second only to Fable 5,” that gap is hard to ignore.
Why it matters
Alibaba has spent the past year building Qwen into China’s default open-weight AI stack, pairing model releases with cloud infrastructure and custom chips. A trillion-plus-parameter multimodal model, if the performance claim holds, would be a real milestone for that strategy.
For enterprise buyers, though, the claim is currently Alibaba’s alone. No independent leaderboard has scored Qwen3.8, and LMArena, where Claude Fable 5 currently holds the top spot, has not yet ranked it. Open weights, when they arrive, will matter more than the press release, because they allow external researchers to actually run the model and compare results.
Businesses evaluating AI tools for AI integration projects should note that preview builds with no model card and no benchmark data are genuinely hard to evaluate. You cannot make a procurement decision or an architecture choice from a claim on X.
Our take
The pattern here is worth naming. Alibaba has a solid track record of following up previews with real open-weight releases and honest benchmark results, as Qwen3.7-Max showed. So the “second only to Fable 5” line is probably a marketing move ahead of a fuller release, not a permanent dodge.
That said, comparing yourself to the current top-ranked model without a single published score is aggressive. It sets an expectation Alibaba will need to meet in public. When the weights and benchmarks land, they will get scrutinized hard, especially given the Kimi K3 competition that just landed three days earlier.
If you are building on Qwen models today, keep watching the open-weights release. The 10% trial pricing through Qoder and QoderWork is a low-cost way to experiment, but we would not commit any production workload to a preview model with no model card.
What to do about it
- Watch the official Qwen release channels for the open-weights drop and license terms before planning any integration.
- Check LMArena and the Artificial Analysis Intelligence Index once Qwen3.8 is officially scored to validate the “second only to Fable 5” claim independently.
- Use the 10% trial pricing on Qoder or QoderWork to run your own prompts on current tasks, but treat results as indicative only.
- Compare against Kimi K3 (2.8 trillion parameters, already released) if multimodal at scale is your actual requirement.
Wait for the benchmarks before making any infrastructure or vendor decisions based on this announcement.
Frequently asked questions
What is Qwen3.8-Max-Preview?
It is Alibaba's latest AI model preview, announced on July 19, 2026, at the World AI Conference in Shanghai. It has 2.4 trillion parameters and can process text, images, video, and documents.
Is Qwen3.8 really better than GPT and Gemini?
Alibaba claims Qwen3.8-Max ranks second only to Anthropic's Claude Fable 5, but has published no benchmarks. No independent leaderboard has scored it yet, so the claim is unverified.
When will Qwen3.8 open weights be released?
Alibaba has promised open weights 'soon' but has not set a specific date or published license terms as of the July 19, 2026 announcement.
How does Qwen3.8 compare to Kimi K3?
Kimi K3, from Beijing Moonshot AI, was released three days before Qwen3.8-Max-Preview and has 2.8 trillion parameters. Both are in the multi-trillion-parameter tier, but neither has been comprehensively benchmarked head-to-head yet.
