Alibaba Qwen Releases Qwen3.8-Max: A 2.4 Trillion Parameter MoE Model and the Most Capable One in the Qwen Family to Date
MarkTechPost Asif Razzaq ● Covered by 2 sources
Alibaba's Qwen team launched Qwen3.8-Max, a 2.4-trillion-parameter model, via API today. Open weights land next week, but a smaller 27B version is the one you'll actually run at home.
Alibaba just pushed out Qwen3.8-Max, and the headline number is hard to ignore: 2.4 trillion total parameters in a mixture-of-experts design that takes text, images and video and spits out text. It's live right now through a hosted API that's OpenAI- and DashScope-compatible, meaning most shops can swap it in with little more than a base-URL change. The open weights are a separate story entirely — those ship next week, alongside a much smaller sibling, Qwen3.8-27B.
That gap between "available" and "deployable" is the real story here. At 2.4T parameters, the Max checkpoint is a multi-node datacenter object, not something you load onto a workstation. Alibaba hasn't said how many of those parameters actually activate per token, so nobody outside the company can model serving costs yet. The 27B model is the one that fits on ordinary on-prem GPUs, which makes it the practical choice for any team that isn't running its own cluster.
The spec sheet is genuinely large. A 1M-token context window, with max input topping out near 991K tokens and output capped at 131K. Pricing lands at $2 per million input tokens and $6 per million output, and cached input runs eight times cheaper than fresh input — so keeping a stable prompt prefix matters more than trimming prompt length. Five built-in tools ship with the Responses API, including code interpretation and web search, aimed squarely at coding agents, legal document review, and long-video indexing work.
Benchmarks tell a mixed but interesting story. Qwen3.8-Max beats Claude Opus 4.8 on Terminal-Bench 2.1 but trails GPT-5.6 Sol, and it falls well behind Fable 5 on SWE-bench Pro and FrontierSWE. Where it clearly shines is multimodal and agentic work — OSWorld-Verified, CAD benchmarking, document parsing — and the year-over-year jump versus Qwen3.7-Max is dramatic in places, like FrontierSWE nearly doubling. Raw reasoning barely moved, though; GPQA Diamond ticked up just two-tenths of a point.
Two details deserve a skeptical eyebrow. The multimodal comparisons use Qwen3.7-Plus as the baseline rather than the actual predecessor Max model, which inflates the apparent leap. And Alibaba's own reinforcement-learning scaling curve peaks around 4,000 training environments before actually declining — not the clean upward line vendors usually show off.
My take
I like that Alibaba is shipping open weights at all, but let's not pretend a 2.4T-parameter checkpoint with no disclosed activation count and no license yet is "open" in any meaningful sense for 99% of developers — the 27B model is the one that matters, and it's the one getting buried under the headline number. This is the same trick every lab pulls now: lead with the flagship's benchmark wins, quietly bury the caveat that your own scaling curve peaked and then declined.
Read more about this at: MarkTechPost
