Artificial Intelligence — Open-Weight Frontier Models

Alibaba Unleashes Qwen 3.8 — A 2.4-Trillion-Parameter Model Built to Chase Fable 5

Alibaba has launched a 2.4-trillion-parameter flagship, Qwen 3.8, with open weights promised to follow — and the launch doubles as a quietly competitive sibling rivalry with the lab it owns a stake in.

• The new Qwen 3.8 Max Preview reaches 2.4 trillion parameters, only about three months after Qwen 3.6 Max Preview.
• Alibaba calls it second only to Anthropic's Fable 5, and it is already live on its cloud's token-based pricing, alongside the coding-oriented Qoder and QoderWork workspaces.
• The model ships as a closed preview first, with open weights promised to follow — the same staged release pattern Alibaba has used as its models have grown.

The headline number is what gets shared, but the more interesting story is the speed of the climb. Alibaba's first trillion-parameter model appeared in September 2025. By April 2026 it had a second, larger preview, and now, less than a year later, it has stepped to 2.4 trillion. That places Qwen 3.8 in the same scale bracket as Moonshot AI's Kimi K3, which launched at 2.8 trillion and briefly held the title of the largest open-weight model ever released.

The comparison is not accidental. Alibaba owns roughly 36% of Moonshot AI. Qwen 3.8's launch is therefore as much an in-house rivalry as it is a global one — sibling labs in Beijing and Hangzhou quietly measuring themselves against each other, while both also compete with z.ai's GLM 5.2 on the broader Chinese frontier and with OpenAI and Anthropic overseas. For anyone tracking where capability is heading, the practical point is that 2-trillion-plus models are no longer singular, once-in-a-cycle events. They are a tier.

What makes Qwen 3.8 notable beyond raw size is the release pattern. Rather than publishing the full weights on day one, Alibaba launched a proprietary preview first and promised open weights later. That staged rollout lets the company validate performance at scale before handing the weights to the community — a move increasingly common as models push past the point where one bad early release is easy to patch. For the open-source ecosystem the takeaway is familiar: the weights eventually arrive, but the first few weeks belong to the cloud that hosts them.