Biplov GautamKev appeared on 19 September 2026 from Jared Palmer, VP of Engineering at Cognition, and it has become the reference open implementation for one reason: it publishes the numbers most people would rather not. Every checkpoint ships with its accuracy on a locked out-of-domain test, its expected calibration error, and a fitted temperature.
The design
Take a Qwen base — Qwen3.5-4B-Base for the reference model — and freeze it. Train a rank-16 LoRA adapter (33.8M parameters) and a pointer head that, given the state and the typed questions in one forward pass, returns a distribution over each question's options. No text is generated; the head points at answers.
The repository holds only the adapter and the head, so kev-4b is a 130 MB download plus a 5 MB head, and the base comes from Qwen. That is also why the family scales so easily: the same recipe on Qwen3.5-0.8B-Base gives kev-0.8b, on Qwen3.5-9B-Base kev-9b, and on Qwen3.8-27B kev-27b.
The numbers
On the locked transfer-v4 test, as served with the shipped temperature of 2.41:
| Model | Out-of-domain accuracy | Calibration error |
|---|---|---|
| kev-0.8b | 69.7% | — |
| kev-4b | 83.8% | 0.042 |
| kev-9b | 85.2% | Brier 0.237 |
In-distribution, kev-4b scores 87.3%. No Jev outputs were used in training. Model time is 36–179 ms on an L40S; Apple silicon is slow, about 495 ms per sample on an M5, because the DeltaNet kernels are not tuned for it.
Running it
Serving context is 8,192 tokens (trained to 384 state tokens), Choice and Score
take up to 255 options, and the server implements TypeSafe's /v1/systemone
contract exactly, so the typesafe-sdk works against Kev by changing the
base URL.
pip install systemonemodels
systemone pull jared-palmer/kev
The pull gives you the adapter, the head, the tokenizer and the training and provenance records; Kev's own runtime loads the base from Qwen. Kev's page on the registry records every reported number with the suite it came from, and links the sibling sizes.
Where it fits
If you want Jev's API shape with weights you own, on a GPU, Kev is the straightforward choice: Apache-2.0, three sizes, and calibration you can read before you download. If you need CPU inference, look at the encoder families — Laya and Von — and if you need something that scores against a large candidate set, CLM.