27B dense language model; 27,781,427,952 total BF16 parameters including vision encoder
262,144 tokens native; extensible to 1,000,000 tokens with YaRN
Apache License 2.0
Compact open-weight dense model from the Qwen3.8 generation, released as a native vision-language model supporting text, image, and video input with text output. Qwen3.8-27B builds on the Qwen3.5 hybrid architecture with a 5,120-dimensional hidden state and 64 language-model layers arranged as 16 repetitions of three Gated DeltaNet-plus-FFN blocks followed by one gated-attention-plus-FFN block, supplemented by Multi-Token Prediction training. The model supports 262,144 tokens natively and extension to one million tokens through YaRN. Unlike the open Qwen3.8-2.4T-A95B checkpoint, the 27B release retains native vision and video understanding and permits thinking to be disabled per request. Thinking is enabled by default, with xhigh, medium, and low reasoning-effort levels, while preserve_thinking can retain reasoning traces across multi-turn agent workflows. Qwen positions the model for coding, professional work, research, long-horizon agents, autonomous planning, environment-feedback handling, computer interaction, and broad agent-harness compatibility. Published evaluations report 73.0 on Terminal Bench 2.1, 61.7 on SWE-bench Pro, 42.2 on DeepSWE 1.1, 79.0 on QwenSWEBench, 70.7 on CoWorkBench, 79.5 on IFBench, and 89.2 on GPQA Diamond. The approximately 55.6 GB BF16 release is distributed across 18 safetensors shards with official deployment guidance for Transformers, vLLM, SGLang, and TokenSpeed.