Qwen3.8 2.4T A95B (Max) represents Alibaba's first open-weight release of a Max-class model, distributed under Apache 2.0 so enterprises and research labs can self-host a roughly 2.4-trillion-parameter mixture-of-experts system rather than relying solely on an API. The model shares the Qwen3.8 family's Hybrid Gated DeltaNet architecture, in which three of every four attention sublayers use linear attention and only one retains standard quadratic attention, a design that supports a native ~262K context window without the memory cost of pure quadratic attention. This combination of openly downloadable Max-tier weights and an efficiency-oriented attention pattern marks a meaningful step toward making frontier-scale reasoning more practical for teams running their own infrastructure.

The model is built for demanding agentic and multimodal workloads: it accepts text, image, video, and PDF inputs and returns text, ships with thinking mode enabled by default, and exposes tunable reasoning effort plus a preserve-thinking flag to maintain coherent multi-turn agent behavior. The same Qwen3.8 generation has demonstrated clear agentic gains over the prior API-only line, with the 27B sibling outperforming its predecessor on SWE-bench Pro (61.7 vs 57.6) and CoWorkBench (70.7 vs 65.1), suggesting the 2.4T-A95B Max inherits a similarly strong agentic coding and office-task foundation. Practical fit centers on cloud or large-cluster deployments that need Max-class reasoning, long-context analysis of mixed-media documents, and tool-using workflows where open weights and adjustable reasoning depth matter more than running on a single workstation.