Alibaba Previews Qwen3.8-Max, a 2.4-Trillion-Parameter Model It Says Trails Only Fable 5
The multimodal Mixture-of-Experts model is Alibaba's first Qwen release past the trillion-parameter mark to handle text, images, video, and documents, though no independent benchmarks back the claim yet.
Alibaba’s Qwen team unveiled a preview of Qwen3.8-Max on July 19, a 2.4-trillion-parameter multimodal model the company claims ranks just below Anthropic’s Fable 5 on internal evaluations — ahead of every other lab, Chinese or Western.
It’s a sparse Mixture-of-Experts design, so only a fraction of those 2.4 trillion parameters activate per token, keeping inference costs manageable despite the size. More notably, it’s the first Qwen model past the one-trillion-parameter mark to go multimodal, handling text, images, video, and document understanding in a single architecture rather than bolting vision onto a text-only base. Alibaba says it should beat predecessor Qwen3.7-Max specifically on coding and “complex productivity” work — full-stack development, data analysis, office-style tasks — the exact territory where enterprises decide whether to switch models.
Take the ranking claim with real caution. Alibaba published it without benchmark numbers, license terms, or technical documentation. No third party has run Qwen3.8-Max against Fable 5, GPT, or Gemini yet. “Second only to Fable 5” is a marketing line until someone outside Alibaba reproduces it.
What is verifiable: the preview is live now, priced at 10% of standard rates through Alibaba’s Token Plan, Qoder, and QoderWork platforms. The full release is planned as open-weight, continuing Qwen’s pattern of shipping frontier-scale models without a license paywall — a strategy that’s steadily eroded the pricing power of closed-weight competitors over the past year. DeepSeek and Moonshot’s Kimi K3 have run the same playbook this month; Qwen3.8-Max lands just days after Kimi K3’s open-weight launch, in what’s becoming a genuine three-way race among Chinese labs to define the open-weight frontier.
For teams evaluating models, the practical takeaway is patience. Alibaba’s internal numbers are directionally useful but not something to build a procurement decision on. Wait for the open-weight release and the independent benchmarks that follow — likely within weeks, given how fast the community re-tests every major Qwen drop.