Alibaba Says Its New Qwen Model Trails Only Anthropic’s Fable 5

Alibaba previewed its newest flagship AI model this weekend and called it second only to Anthropic’s Fable 5 among AI systems worldwide. Qwen3.8-Max-Preview packs 2.4 trillion parameters, handles multiple data types, and is already live on Alibaba’s coding platforms. What it did not arrive with: a benchmark table, a license, or any word on how many of those parameters actually switch on for a given query.

The preview landed three days after Moonshot AI, a Beijing startup that counts Alibaba among its own backers according to VentureBeat, shipped a bigger 2.8 trillion-parameter model called Kimi K3. The timing points to a contest closer to home: Alibaba racing a company it helped fund, with Anthropic’s newest model used as the measuring stick for both sides.

A 2.4 Trillion-Parameter Model Ships Without Its Receipts

Alibaba Group, the Hangzhou-based e-commerce and cloud company, posted the preview on X over the weekend. Its own language put Qwen3.8-Max-Preview behind just one system in the world: “second only to Fable 5,” the newest model line out of Anthropic. Alibaba’s Hong Kong-listed shares (9988.HK) rose 3.73% around the announcement, and the company described the model as among the most powerful available today.

What showed up alongside that claim was thinner. MarkTechPost, an AI research news outlet, reported that Qwen3.8-Max-Preview shipped with no benchmark table, no license, and no disclosed active-parameter count, the smaller figure that determines real-world speed and cost in a mixture-of-experts model like this one.

What We Know:

• Qwen3.8-Max-Preview carries 2.4 trillion total parameters and is multimodal.
• It is live now on Alibaba’s coding platforms and was announced in a post on X.
• Alibaba’s own framing calls it “second only to Fable 5” among the systems it tested.

What’s Unconfirmed:

• No benchmark scores have been published for the preview.
• No license terms have been released, despite Alibaba’s history of open-weighting earlier Qwen models.
• The active-parameter count, which shapes real-world cost and latency, has not been disclosed.

That gap between claim and proof came three days after a bigger model from much closer to home.

Moonshot’s Bigger Model Landed First

Moonshot AI, the Beijing startup behind the Kimi chatbot, released Kimi K3 on July 16, three days before Alibaba’s preview went up. At 2.8 trillion parameters, Kimi K3 is larger than Qwen3.8-Max-Preview and, per outlets tracking the launch, the largest open-weight model released to date. It runs a 1-million-token context window and introduces a new attention mechanism the company calls Kimi Delta Attention.

Moonshot has not fully delivered on its own promises either. A technical breakdown of the release notes that full weights are not due until the open-source community’s July 27 deadline, more than a week after the model went live through Moonshot’s apps and API.

VentureBeat described the launch as a comeback for a company whose market position had eroded for 18 months after a rival Chinese lab, DeepSeek, surged ahead. Independent testers reportedly found Kimi K3 beating every other model, open or closed, in at least one high-profile benchmark arena, a result Alibaba’s own “second only to Fable 5” claim would need to top.

Why Is Alibaba Racing a Company It Funds?

Alibaba is one of Moonshot AI’s backers, according to VentureBeat, which makes Qwen3.8-Max-Preview’s rushed timing unusual: Alibaba effectively raced past a startup it helped finance days after that startup’s own model went live, competing for the same enterprise developers and the same claim to China’s AI crown.

Four parties have a stake in how this plays out, and not all of them benefit equally:

• Alibaba gets a headline claim of near-frontier performance and a product to sell through its coding platforms.
• Moonshot AI gets outflanked in its own launch week by a company that helps finance it.
• Developers get two large, open-weight-leaning Chinese models inside a single week, each promising frontier-adjacent performance for little or no cost.
• Anthropic gets used as the yardstick for both launches without weighing in on either claim.

Alibaba’s arrangement is not unusual for Chinese tech. Tencent has run a similar playbook, backing Kling AI through a round valuing the video-AI startup at $18 billion, even as Kling competes for some of the same enterprise budgets Tencent’s own cloud business chases. Big Chinese platforms are increasingly funding the very labs now racing to out-ship them.

The Price of Second Place, According to Anthropic

Fable 5 is the model Alibaba chose as its yardstick, and it is worth knowing what that yardstick actually costs. Anthropic released Claude Fable 5 on June 9 as the first entry in a new “Mythos-class” tier sitting above its existing Opus line. Anthropic’s own announcement called it a clear step forward on agentic coding, and independent SWE-Bench Pro testing puts it around 80%, against roughly 69% for Opus 4.8.

None of that comes cheap or unrestricted. Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, twice the cost of Opus 4.8. It is cloud-only: there are no open weights to download and no way to run it outside Anthropic’s infrastructure or its cloud partners. A companion model, Mythos 5, shares the same underlying capability but skips Fable 5’s safety classifiers and remains limited to a small release program.

Fable 5 also spent part of its short life unavailable outside the United States. Export controls on the model were lifted June 30, with global access restored the next day. That timeline means Alibaba measured its own preview against a model that had itself been under U.S. export restriction just three weeks earlier.

Laid out together, the numbers show why Alibaba reached for a Silicon Valley name rather than its neighbor in Beijing. Kimi K3 is the bigger model by parameter count and shipped first.

Alibaba Has Run This Play Before

This is not the first time Alibaba has previewed a Qwen model with a bold claim attached, and its own record makes the current silence more notable. The original Qwen3 launched in 2025, topping benchmarks such as AIME25 and LiveCodeBench among open models of its size. Qwen3-Max followed with a reasoning variant Alibaba said was posting perfect scores on tough math benchmarks.

The cadence has only sped up since. Qwen3.5 arrived with claims of beating GPT-5.2, Claude Opus 4.5, and Gemini 3 Pro across most evaluation categories. Qwen3.7-Max reportedly cracked the global top five on one closely watched agent ranking this past June. Qwen3.6-27B, a smaller dense model released in April, posted higher reasoning scores than Alibaba’s own larger models from months earlier.

Each of those releases came bundled with a benchmark table and, in most cases, an open license on day one. Qwen3.8-Max-Preview, so far, has shipped with neither.

No License, No Benchmarks, Not Yet

Independent trackers are already watching for the numbers Alibaba has not supplied. BenchLM.ai, a site that catalogs published model evaluations, listed zero of the 321 benchmark categories it tracks as complete for Qwen3.8-Max-Preview this week.

Shipping first and disclosing later is not unique to Alibaba right now. xAI recently open-sourced Grok Build’s code while keeping its kill switch, the same kind of gap between the openness a company advertises and the openness it actually hands over.

Alibaba’s claim of ranking second in the world may hold up once the benchmark table, license, and active-parameter count arrive. Until then it is a post on X, measured against a rival’s model that Alibaba helped fund three days before its own preview went live.