⚡️ you don't need a bigger model — or Opus prices — to beat Claude on desktop … · @QwenAIFans · TwKit

@QwenAIFans

⚡️ you don't need a bigger model — or Opus prices — to beat Claude on desktop tasks Simular's Agent S4 left Qwen3.8-27B's weights untouched. It evolved the Python "shadow interfaces" the agent uses instead. Same 27B, OSWorld V2 (108 tasks, strict binary): • fixed interfaces: 5.56% • evolved interfaces: 21.3% (3.8×) • beats GPT-5.5 13.0% and Opus 4.8 20.6% • ~$1.10/task vs $25.50 / $72.40 Caveat: V2 only counts full pass (no partial credit), so the % looks low — under that rule it's open-source SOTA. via Simular 👇 https://www.simular.ai/articles/agent-s4-evolving-the-interfaces-computer-use-agents-work-throughagen

曝光 246 · 评论 2 · 点赞 4 · 书签 1