1.0
New
Kerrfield One is generally available
The full model is now available to everyone on the API and in the app, with a 142-page system card published the same day.
0.9.6
Improved
Thinking budgets up to 30 minutes
Team and Enterprise requests can now think for up to 30 minutes. A request never spends more than its budget.
0.9.5
New
Stream every reasoning step
The API now streams thinking.step, thinking.branch and thinking.check events, so your UI can show progress as it happens.
0.9.4
New
Open 8B distill under Apache 2.0
Kerrfield One Distill is released with weights, recipe and a 4-bit laptop build.
0.9.3
Improved
PDF and image inputs
Attach PDFs, charts and screenshots. The model reads tables and figures, not just the text layer.
0.9.2
New
Confidence on every answer
Every answer now carries a calibrated confidence score, and you can set a threshold below which it asks instead.
0.9.1
Fixed
Faster first token
First-token latency is down to 280 ms at p50, and long attachments no longer delay the first streamed step.
One endpoint.Streaming by default.
Send a question and whatever it needs to read. Set a thinking budget. Stream every step as it happens, or wait for the final answer and its confidence.