Kev
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.
This entry is kev-4b on Qwen3.5-4B-Base, the family's reference size; the adapter is rank 16 (33.8M parameters) and the repository holds the adapter and head, not the base weights. On the locked out-of-domain test (transfer-v4) as served it scores 83.8% with an expected calibration error of 0.042; in-distribution 87.3%. Siblings, all Apache-2.0: kev-0.8b (69.7% out of domain), kev-9b (85.2%, Brier 0.237), kev-27b on Qwen's post-trained Qwen3.8-27B (89.6% on the same test, the most accurate Kev; 84.8% on development sources, where Jev scores 85.7%), and kev-8b, kev-0.6b and kev-0.5b on older bases. Serving context is 8,192 tokens; Choice and Score take up to 255 options. No Jev outputs were used in training. Model time is 36–179 ms on an L40S; Apple silicon is slow (DeltaNet kernels).
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one calibrated probability
- classify — assigns a category from a fixed taxonomy
- route — sends the input to one of several destinations
At a glance
| Parameters | 4B |
| Base model | qwen/qwen3.5-4b-base |
| Maker | Jared Palmer |
| Released | 2026-09-19 |
| License | apache-2.0 |
| Reported accuracy | 83.8% |
| Reported latency | 36–179 ms model time on L40S; ~495 ms per sample on Apple M5 |
Get the weights
pip install systemonemodels
systemone pull jared-palmer/kev
The files are served from the maker's Hugging Face repository, jaredpalmer/kev-4b, and verified against the checksums recorded here.