Qwen3.8-27B tops Opus 4.6 Max in computer use

// 1h agoBENCHMARK RESULT

Qwen3.8-27B tops Opus 4.6 Max in computer use

New benchmark figures indicate that the 27-billion-parameter open-weight model Qwen3.8-27B substantially outperforms proprietary frontier model Claude Opus 4.6 Max at computer and device control tasks. Evaluated on OSWorld-Verified, Qwen3.8-27B scored 84.3 compared to Opus 4.6 Max's 72.7, while on AndroidWorld it achieved 81.9 versus 62.0. These wide margins demonstrate that targeted open-weight architectures are becoming formidable contenders against the largest closed models in practical GUI grounding, screen comprehension, and automated desktop and mobile operating system navigation.

// ANALYSIS

Smaller open-weight models dominating computer-use benchmarks will rapidly accelerate the shift toward private, locally deployed autonomous agents. Leading by 11.6 points on OSWorld-Verified and nearly 20 points on AndroidWorld proves open-weight models can excel at complex, multi-step interface workflows. A 27B parameter model can also be served locally on accessible consumer or workstation hardware, eliminating latency and privacy concerns tied to streaming screen captures to closed APIs. Furthermore, superior performance in agentic execution by open weights reduces reliance on costly closed-source subscriptions for OS automation.

// TAGS

qwenopen-weightbenchmarksosworldandroidworldcomputer-useagent

DISCOVERED

1h ago

2026-09-12

PUBLISHED

1h ago

2026-09-12

RELEVANCE

8/ 10

AUTHOR

Oluwaphilemon1

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

RESEARCH1h ago

26-week simulation reveals LLM town economy stalls

In a paper titled "But How Would AI Agents Run a Town's Economy?", researchers placed 100 memory-equipped LLM agents in charge of a closed, money-conserving spatial economy mapped to Pokhara Lakeside geography for up to 26 simulated weeks across 91 validated runs and 2.44 million decisions. The study found that while a 12x tourist demand shock increased business revenue by 4.62x, monetary transmission abruptly stopped there: wages moved by only 1.03x and just 0.3% of 3,981 menu items were ever repriced. Similarly, when agents were given randomized cash transfers, 96.7% of the funds remained unspent hundreds of steps later, demonstrating a near-zero marginal propensity to consume. Crucially, ablation experiments revealed that swapping the underlying LLM altered every measured economic outcome, whereas deleting agents' memory produced no detectable changes.

NEWS1h ago

Project Astra generates ambient music from screen activity

Developer and Instructor library creator Jason Liu (@jxnlco) shared an exploratory use case leveraging Google DeepMind's Project Astra to monitor his computer screen and desktop usage history to synthesize continuous ambient background music. By tapping into Astra's real-time multimodal perception and generative audio capabilities, the workflow transforms on-screen context into an adaptive, personalized soundscape designed for focused computer work.

FUNDING1h ago

OpenAI brings on Git AI team for Codex

OpenAI has welcomed Aidan Cunniffe and Sasha Varlamov from Git AI, an open-source tool for tracking agent code contributions, to bolster repository attribution for Codex. The integration aims to give businesses granular visibility into agent value, while OpenAI commits to maintaining Git AI as an open-source project with continued investment.