Jev passes 8/9 computer-use tasks, makes decisions 13.6x faster than Astra/Codex
iamrobotbear · x · 2026-09-17
francedot tested TypesafeAI's new Jev model for computer use with Cua Driver: it passed 8/9 executable tasks (vs 9/9 for Astra/Codex), with median decisions 13.6x faster and median task time 27% lower.
The author's takeaway: once a model has a usable representation of the UI, much of computer use becomes a general decision problem. Solid showing for a first public release, though the test was small.
More from coding & agent
- Nautilo pitches time-limited delegated E2E encryption and hierarchical memory for AI agents — Dan_Jeffries1 · 2026-09-17
- Why Nautilo rebuilt chat: WhatsApp and Signal were never designed for agents — Dan_Jeffries1 · 2026-09-17
- Nautilo uses memory namespaces so AI agents can't leak what they can't access — Dan_Jeffries1 · 2026-09-17
- Nautilo launches: a 100% open-source org harness with a personal 'Genie' agent for every user — Dan_Jeffries1 · 2026-09-17
- Open tutorial: training a search agent with GRPO, every rollout browsable and code open-sourced — dejavucoder · 2026-09-17
- Unsloth Desktop Launches: Train & Run 500+ Local Models, 70% Less VRAM — danielhanchen · 2026-09-17