Release
9 min
Kerrfield One Distill: an open 8B model
A small reasoning model you can run on one GPU, released under Apache 2.0 with its training recipe.
Kerrfield research
Not every problem needs the full model. Distill is an 8B model trained on 2.1 million reasoning traces from Kerrfield One, and it runs on a single 24 GB GPU.
It keeps 81% of the full model’s score on our maths set at a twentieth of the cost, and it is free to use, change and ship.
What we released
Weights under Apache 2.0
The full distillation recipe and data mix
A 4-bit build for laptops
We will keep improving it in the open. Issues and pull requests are welcome.
Keep reading
More from the lab.
System card
32 min
Kerrfield One system card
How Kerrfield One was trained, what 11,400 hours of red-teaming found, and every mitigation we shipped before launch.
Paper
18 min
Thinking longer, not bigger
A 70B model given four minutes to think beats a model ten times its size answering at once. Here is why, and where it stops working.
Paper
14 min
Calibration you can check
When Kerrfield One says it is 80% sure, it is right about 80% of the time. How we trained for honest confidence, and how to verify it yourself.