Qwen3.8-27B发布：多项Agent评测超Claude Opus 4.6，24GB显卡可本地运行
开源 Qwen3.8-27B 以 270 亿参数在多项评测超过 Claude Opus，且 24GB 显卡可本地运行，降低高端 Agent 使用门槛。
Qwen3.8-27B 是阿里开源的 270 亿参数模型，用消费级显卡就能本地跑。它在软件工程和 Agent 评测中得分超过 Claude Opus 4.6 Max，支持图像、文档和视频输入，原生上下文 262K token，可扩展到 100 万。这让个人开发者有机会在本地运行接近顶级闭源模型的 Agent。
正文摘录
- title: "Source God Launch! A Consumer-Grade GPU Runs an 'Opus-Level' Agent — Qwen3.8-27B Beats Claude on Multiple Benchmarks" - sourcecompany: 量子位 (QbitAI) - bodymarkdown: With a total of 27 billion parameters, its performance across multiple official Benchmark software engineering and Agent evaluations is, frankly, making Claude sweat a little—— On the Agent programming benchmark SWE-bench Pro, the 27B model beat Claude Opus 4.6 Max by a margin of 8.3 points. On QwenSWEBench, which better tests real-world software engineering capability, the lead even widened to 15.2 points: Native multimodal support, a 262K native context window, extensible up to 1 million tokens, with a strong emphasis on Coding, professional work, and long-horizon Agent capabilities — all present and accounted for. T…