CUA-S1 Open-Sourced: 706K-Param On-Device Computer-Use Model Hits 1ms Inference, 99.95% Accuracy
GabGarrett · x · 2026-09-21
trycua open-sourced CUA-S1-FORMS, the first in a family of small specialized "System One" models built for computer use. Converted to Core ML, it runs fully on-device: 706K parameters, 1ms inference on M5 Pro, up to 98.2% of ops on the Apple Neural Engine, and 99.95% accuracy across 24,370 decisions, with int4 quantization also achieved. SDK and model weights are public.
More from coding & agent
- Memory is not permission: four boundaries to keep personal agents from acting on their own — sujingshen · 2026-09-21
- LlamaIndex founder open-sources DocJev: doc classification/splitting 6x faster than GPT-5.6 — llama_index · 2026-09-21
- Dev benchmarks Jev vs a local 7B model for LLM routing: Jev faster, 7B more accurate — tinyfool · 2026-09-21
- Periodic HANDOFF.md: A Practical Trick to Survive Context Compaction in Claude Code — MikePFrank · 2026-09-21
- User Praises Codex for Seamless Migration of Skills and Workflows from Other AI Assistants — shensi · 2026-09-21
- Ant International's Code2Skill mines verifiable agent skills from code at scale — ant-intl · 2026-09-21