Qwen3.8-27B

About

Qwen3.8-27B (Qwen/Qwen3.8-27B) is the dense mid-size multimodal model of Alibaba's Qwen3.8 generation - 27B parameters (dense; FFN-based, no MoE) handling text, image and video input, released August 5, 2026 under Apache 2.0. Its 64-layer hybrid stack repeats 16x (3x (Gated DeltaNet -> FFN) + 1x (Gated Attention -> FFN)): 48 linear-attention layers (Gated DeltaNet, 48 V / 16 QK heads, head dim 128) and 16 full-attention layers (Gated Attention, 24 Q / 4 KV heads, head dim 256, RoPE dim 64), with hidden size 5120, FFN intermediate dimension 17,408, 248K vocabulary, Multi-Token Prediction (MTP) trained with multiple steps, and a 262,144-token native context extensible to 1,000,000.

Qwen3.8 delivers comprehensive improvements across coding, professional work, research and long-horizon agentic tasks; stronger autonomous planning and environment-feedback handling for reliable end-to-end agent execution; broader compatibility with popular harnesses and development tools; flexible thinking control - thinking mode on by default, disable-able per request, reasoning depth tunable via reasoning_effort, and reasoning context retained across messages via preserve_thinking; and native vision-language understanding from STEM diagrams and documents to hour-scale videos.

Training Data Pre-training & Post-training. Trained with multi-token prediction (MTP) steps. Post-trained with Thinking Preservation feature. Flexible thinking control with reasoning_effort levels.