Simular's Agent S4 evolves custom interfaces so a 27B open model beats bigger ones on desktop tasks

Oct 23, 2026 · Source: X

Simular's Agent S4 acts as a meta-agent that studies example tasks and evolves custom ways for a model to operate software, instead of relying on fixed screenshots and clicks. With it, the open Qwen3.8-27B model scored higher on the OSWorld V2 desktop benchmark than much larger closed models, at about $1.10 per task.

Read the original on X ↗

Related on AgentScout

OSWorld

Benchmark for multimodal computer-use agents on open-ended tasks in real desktop environments.

More agent news from Oct 23, 2026

• Moonshot AI releases Kimi K3.1 Pro with stronger coding and agent scores (X)
• Cursor coding agents can now use your computer to run and test the apps they build (X)
• MiniMax open-sources M3.1 Flash, a model built for agentic coding and computer use (X)
• Big banks spot unauthorized AI agents in customer accounts and move to restrict them (X)
• Anthropic previews Claude Duet, a proactive agent app that watches your email and calendar (X)

Newer: MiniMax open-sources M3.1 Flash, a model built for agentic coding and computer use · Older: Big banks spot unauthorized AI agents in customer accounts and move to restrict them

Latest agent news · All October 2026 stories · News archive

Summary written by AgentScout from the original report; follow the source link for the full story.