Alibaba's Qwen3.8 Claims It Trails Only Claude Fable 5

Alibaba says it has built a model that beats everything on the market except Anthropic's flagship. It said so on X, without a benchmark table, a model card, or a number for how many of its 2.4 trillion parameters actually activate per token.

The model is Qwen3.8-Max-Preview, unveiled Sunday at the World AI Conference (WAIC) in Shanghai, and Alibaba's own framing leaves little room for modesty: "one of the most powerful models available today, comparable to leading frontier AI models, second only to [Claude] Fable 5." That claim comes from the company that made the model. Nobody outside Alibaba has tested it yet.

TL;DR

Qwen3.8-Max-Preview is a 2.4 trillion-parameter multimodal MoE model with a 1M-token context window, Alibaba's first to cross the 1-trillion-parameter mark
It handles text, images, video, and documents, and ships with thinking mode, function calling, and built-in tools
Alibaba claims it ranks second only to Claude Fable 5, with no independent benchmarks published to support that
Preview access is live now via Token Plan, Qoder, and QoderWork at 10% of standard pricing; open weights are promised "soon" with no date
The launch lands three days after Moonshot AI's open-weight Kimi K3, turning this into a direct answer in China's model race

What Alibaba Announced

Qwen3.8-Max-Preview is a sparse mixture-of-experts model, meaning only a fraction of its 2.4 trillion parameters fire for any given token. Alibaba hasn't said how large that active slice is. Total parameter count is the one number the company was willing to share, alongside a 1-million-token context window and support for thinking mode, function calling, and built-in tool use.

It's also Alibaba's first model past the trillion-parameter line, and its first multimodal system at that scale, reading and reasoning over images, video, and documents rather than text alone. Qwen developer Shuai Bai called it out as a milestone in a post accompanying the launch, describing it as the team's first multimodal model above 1 trillion parameters.

Model	Total params	Active params	Context	Open weights
Qwen3.8-Max-Preview	2.4T	Not disclosed	1M tokens	Promised, no date
Kimi K3	2.8T	Not disclosed	Not disclosed	Yes, at launch
Qwen3.7-Max	Not disclosed	Not disclosed	1M tokens	No
Claude Fable 5	Not disclosed	N/A	Not disclosed	No

The Pricing Play

Preview access runs through Alibaba's Token Plan subscription, plus the Qoder and QoderWork agentic coding platforms, all priced at 10% of what the eventual standard rate will be. Alibaba hasn't said what that standard rate is, so the discount is a promise with no baseline attached. Anyone building against the preview today is pricing a product Alibaba could still change before it ships.

The Open-Weight Promise

A fully open-weight release of Qwen3.8 is coming "soon," per Alibaba, with no committed date. Qwen3.7-Max shipped closed-weight in May, part of a pattern where Alibaba reserves its biggest models for paid access while open-sourcing smaller variants. A preview gated behind a subscription isn't the open-source release Alibaba is promising. It's a teaser for one.

Qwen is Alibaba Cloud's LLM family, spanning models from 0.6 billion to now 2.4 trillion parameters. Source: commons.wikimedia.org

What It Does Not Tell You

The gap between what Alibaba announced and what it proved is the story here. No benchmark scores. No model card. No independent evaluation from a third party. No active-parameter count, which is the single most useful number for judging how expensive this model actually is to run compared to a dense 2.4-trillion-parameter system. "Second only to Fable 5" is a ranking Alibaba assigned itself, using benchmarks it hasn't shown anyone.

Compare that with how Qwen3.7-Max launched in May: a published model card, Terminal-Bench 2.0 and SWE-Bench Pro scores, and pricing at $2.50 per million input tokens and $7.50 per million output tokens, all available on day one. Qwen3.8-Max-Preview shipped with none of that. The preview is live, but the receipts are not.

The Kimi K3 Timing

Three days separate this announcement from Moonshot AI's Kimi K3 launch, a 2.8-trillion-parameter model that shipped open-weight from day one and immediately topped the Frontend Arena leaderboard before tripling its API price. Alibaba didn't mention Kimi K3 by name in its announcement. It didn't need to. A company previewing a closed model days after a rival ships an open one at similar scale is making a competitive statement whether it says so or not, and our own comparison of Kimi K3 against Claude Fable 5 is a reasonable proxy for where Qwen3.8 would need to land to back up its claim.

Alibaba on display at the World AI Conference in Shanghai Alibaba has used WAIC as its venue of choice for major model shows two years running. Source: commons.wikimedia.org

Why WAIC Matters Here

Alibaba didn't pick a random week for this. WAIC is China's flagship AI showcase, running through July 20 in Shanghai with more than 1,400 speakers and over 260 foundation models on display, according to conference organizers. Announcing a 2.4-trillion-parameter model at the country's biggest AI stage, days after a domestic rival's open-weight launch, is as much a signal to Beijing and to Chinese enterprise buyers as it's to developers deciding which API to call.

That context matters more than the parameter count. Qwen3.7-Max closed the gap between claim and evidence with published data. Qwen3.8-Max-Preview, for now, has not.

A 2.4-trillion-parameter multimodal model with a 1-million-token context window is a genuine engineering achievement regardless of where it ranks. But "second only to Fable 5" is a marketing line until Alibaba publishes the benchmarks to defend it, and until the open-weight release actually arrives, developers assessing Qwen3.8 are evaluating a preview, a price promise, and a claim, not a finished product.

Sources: