Qwen 27B Surpasses Claude Opus in Some Coding Benchmarks

The 27B-level open AI model from Qwen has outperformed Anthropic's Claude Opus 4.6 Max in several coding benchmarks. Qwen 3.8-27B scored 61.7 points on the SWE-Bench Pro, surpassing Claude's 53.4 points, and achieved a score of 79.0 on QwenSWEBench. However, in TerminalBench 2.1, Claude maintained an advantage with a score of 78.2. In the repository-level code generation evaluation, NL2Repo-Bench, Claude also exceeded Qwen with a score of 47.6 compared to Qwen's 42.3. Qwen 3.8-27B offers a base context length of 262,144 tokens and an expandable length of up to 1 million tokens, including capabilities for image and video understanding. According to the Hugging Face model card, Qwen 3.8-27B also provides a quantization model path that can connect to local execution tools, which is advantageous for developers looking to operate the model directly on enterprise cloud APIs and personal workstations. However, the feasibility of running it on consumer GPUs may vary depending on the setup. Qwen Cloud introduced Qwen 3.8-Max as a 2.4 trillion parameter MoE model, describing it as focused on coding and specialized tasks. This release highlights a trend where Chinese open models are being compared to American models in coding and agent tasks.