CUA-S1-FORMS: a 2.8 MB open model fills forms locally in 7ms
Cua's first 'System One' model scores typed UI actions in one forward pass — no LLM, no API. Plus stage-only npm tokens and Terminal-Bench on Vercel Sandbox.
Copy markdown
706k parameters, 2.8 MB, MIT-licensed
CUA-S1-FORMS decides fill, check, click, or skip on each form field by scoring typed options in a single forward pass — no text generation, no screenshots. The whole checkpoint is 2.8 MB and trains in under 30 minutes on synthetic data, so you can fine-tune your own specialist.
7-9ms local vs 260-280ms hosted
Scoring runs on your machine in 7-9ms instead of round-tripping to an API at 260-280ms, and it hit 99.7% on the full decision set against 83.6% for TypeSafe's hosted Jev. The honest caveat: it was fine-tuned for this one task on synthetic data and can only pick from pre-extracted values, so read the gap as scoped specialization, not a general win.
Elsewhere: GitHub ships stage-only npm tokens
A new granular token can stage a release for review but can't publish directly: your CI runs `npm stage publish` and a maintainer approves it with 2FA, so a leaked token can't push malware to your public package. It's the migration path before npm ends bypass-2FA publishing in January 2027; needs npm CLI 11.15+ and Node 22.14+.
Elsewhere: Terminal-Bench now runs on Vercel Sandbox
Pass `--env vercel` to `harbor run` and each trial spins up its own Firecracker microVM, so you can parallelize Terminal-Bench, SWE-bench, or OSWorld agent evals (try `--n-concurrent 8`) and swap models through AI Gateway with a single `--model` flag. Needs Harbor 0.22.0+.