Alibaba releases Qwen3.8 open weights under Apache 2.0, up to 2.4T parameters
Alibaba's Qwen team put Qwen3.8-27B and the far larger Qwen3.8-2.4T-A95B on Hugging Face and ModelScope under Apache 2.0, with the 27B dense multimodal model outscoring the bigger Qwen3.7-Plus on coding and office tasks. That packs images, multi-hour video, and 262K native context — one million via YaRN — into weights anyone can self-host, and pushes the open-weights frontier further onto Alibaba's terms. A hosted million-token version is coming to Qwen Cloud.
Source: the-decoder.com ↗
The core model, Qwen3.8-27B, is a multimodal dense model with 27 billion parameters that, according to Qwen, outperforms the larger Qwen3.7-Plus in coding and office tasks.
Alibaba's Qwen team
Why this matters
- → Small open model outperforms larger closed predecessor on key tasks
- → Apache 2.0 weights enable self-hosted multimodal inference at million-token scale
- → Shifts open-weights frontier toward Alibaba's model architecture and licensing terms
Open weights, closed gap