Berget AI's System One model for gating agent commands: a LoRA adapter and a fine-tuned joint schema head on Cloudflare's Clef-Flash that answer noul, choice and score questions over a state in one forward pass. Trained on Swedish and English operations decisions.
An open family of System One models. A LoRA adapter plus a pointer head on a frozen Qwen base returns a distribution per typed question in one forward pass, serves TypeSafe's /v1/systemone contract, and ships a fitted temperature with every checkpoint.
Decides
noul, choice, score
choice, score, noul, classify, route
Architecture
clef
kev
Fine-tuned from
cloudflare/clef-flash
qwen/qwen3.5-4b-base
License
apache-2.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
97.0%
83.8%
Calibration error
—
0.042
Valid action rate
—
—
Median latency
—
—
p95 latency
—
—
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between bev and kev?
bev is from Berget AI and kev from Jared Palmer. Both have open weights you can download and run. Both answer noul, choice and score questions. Only kev answers classify and route. kev is the smaller model, at 4.0B parameters to 9.0B.
Which is more accurate, bev or kev?
They report on different suites — bev 97.0% on Berget held-out risk split (16,902 questions; same operations-traffic corpora as training), kev 83.8% on transfer-v4 (locked, out of domain) — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, bev or kev?
bev: Free (open weights). kev: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run bev or kev locally?
Yes, both: systemone pull berget-ai/bev and systemone pull jared-palmer/kev download the weights.
Evaluation suite
Berget held-out risk split (16,902 questions; same operations-traffic corpora as training)
