Alibaba Opens Qwen3.8 Weights: 27B Multimodal Model Plus a 2.4T Max-Tier MoE

Alibaba’s Qwen team has published open weights for its Qwen3.8 lineup under a permissive Apache 2.0 license, headlined by Qwen3.8-27B. It’s a dense, natively multimodal model that the team positions as small enough to run locally while claiming it beats the larger Qwen3.7-Plus overall, with particular strength in coding and office-style workflows. It ships with a 262K-token native context window that Qwen says can be stretched toward 1M tokens using YaRN scaling.

Alongside the compact model, Qwen also released weights for Qwen3.8-2.4T-A95B, a Max-tier mixture-of-experts model (2.4 trillion total parameters, roughly 95B active) aimed at heavier agent-building workloads. The pitch splits cleanly by use case: the 27B for lightweight, on-device applications and the 2.4T MoE for more demanding agentic systems. Both are distributed through Hugging Face and ModelScope.

The significance is less about any single benchmark—these figures are vendor-reported and unverified—and more about the continued push to put frontier-adjacent, top-tier models into fully open, commercially usable hands. Releasing a Max-level flagship under Apache 2.0 is aggressive, lowering the cost and licensing friction for teams that want to self-host capable multimodal and agent models rather than rent them via API.