Qwen3.8

Alibaba's August 2026 generation spans the widest range of any open family. At the top, Qwen3.8-2.4T-A95B, the open-weight release of the hosted Qwen3.8-Max, is second in size only to Kimi K3 among open models: 2.4 trillion total parameters with 95B active, built from a hybrid stack that interleaves Gated DeltaNet linear-attention blocks with gated full attention, routing through 512 experts per MoE layer. It runs exclusively in thinking mode, handles text only, reads 262K tokens natively with extension toward 1M, and downloads from Hugging Face under the custom Qwen3.8-Max license. At the other end, Qwen3.8-27B arrived in mid-August 2026 under Apache 2.0 as a dense vision-language model that reads images and video and posts strong coding scores, with 4-bit quantized builds around 17GB that fit a single consumer GPU. Recommended serving for the big checkpoints is SGLang, vLLM, or TokenSpeed on multi-GPU clusters, with Alibaba Cloud offering hosted endpoints; the 27B's permissive terms make it the practical choice for commercial local deployment.