Qwen3.8-27B Open-Sourced: Runs 'Opus-Level' Agent on a Consumer GPU, Beats Claude in Multiple Benchmarks

Alibaba has released the open-source Qwen3.8-27B model with 27 billion parameters. On official benchmarks, it surpasses Claude Opus 4.6 Max in software engineering and agent tasks, such as leading by 8.3 points on SWE-bench Pro and by 15.2 points on QwenSWEBench. In agent evaluations, it scores 70.7 on CoWorkBench versus Opus's 68.2, and in multimodal computer and phone operation tests, it achieves 84.3 on OSWorld-Verified and 81.9 on AndroidWorld, compared to Opus's 72.7 and 62.0. The model supports native multimodal understanding, 262K token native context expandable to 1 million, and includes a reasoning effort adjustment (xhigh, medium, low) and preserve_thinking for agent tasks. Its architecture uses 48 layers of Gated DeltaNet linear attention and 16 layers of full attention, enabling long context. It is compatible with Transformers, vLLM, SGLang, and TokenSpeed, and quantized versions are available on Hugging Face, allowing deployment on consumer GPUs like RTX 3090/4090 or Apple Macs. Community tests demonstrate its coding and multimodal capabilities, including high-concurrency stability on NVIDIA GH200 and processing long videos.