Kev model explained: sizes, licence and Jev vs Kev

The Hugging Face repositories belong to the user jaredpalmer; the Kev 27B model card names Jared Palmer as developer and the GitHub repository as the code source. The GitHub repository is licensed Apache-2.0 and was created on 17 September 2026. The 27B and 9B cards date version 2 (“Kev 1.0”) to 30 September 2026; the repositories were created between 18 and 24 September 2026 (Hugging Face API).

All seven repositories state license: apache-2.0 in their metadata. “LoRA” means, per the cards, an adapter plus a pointer head on a frozen Qwen base model; Kev 27B is a fine-tune of all weights. The Kev 8B card calls that checkpoint a previous (Qwen3) generation. This is not our own legal review; training datasets and base models carry their own licences, which the cards list.

Author-reported, not independently verified: per the card, Kev reads one document and typed questions (choice, score, noul) and returns a calibrated probability distribution in a single pass, without generating text. The card describes an implementation of TypeSafe’s public System One API, so the TypeSafe SDK is meant to work unchanged. For Kev 27B the card reports accuracy of 0.8887 on a 656-question test split it describes as out of domain. The card also includes its own comparisons with Jev; we do not repeat them.

Source: the Kev 27B system page and other Kev sizes plus Jev 1.13.0 and Jev 1.14.0 on Benchmark Heaven, read 29 October 2026. Shown is data release v1.7.21 (published 29 October 2026), method v1.7 with the 10-second rule, rescored from the recorded decision times. Capability is the mean of Intelligence and Calibration; it is not accuracy in percent. The individual axes are on the linked pages. Benchmark Heaven has no dedicated Jev-vs-Kev page at the moment, so the values come from the individual system pages.

Kev was measured as self-hosted weights on the evaluator’s own hardware, Jev through TypeSafe’s hosted API; runs span 2 to 16 October 2026 across different measurement releases. The “kev 4B” row scores the Qwen3 variant; the current Hugging Face checkpoint uses Qwen3.5-4B-Base. No Benchmark Heaven page was available for Kev 0.8B. These are different runs and access routes, not same-workload, speed or price comparisons, and not a ranking.

Self-host: the model card says to serve Kev 27B with the Kev repository on an 80 GB-class GPU; the weights are 51 GB and, per the card, the longest inputs need more than 80 GB. The two lines below are taken from the card. We have not run them and have not downloaded any weights. The smaller sizes are documented in their own repositories.

git clone https://github.com/jaredpalmer/kev.git && cd kev && uv sync --extra serve
uv run --extra serve python -m kev.serve --run jaredpalmer/kev-27b --port 8008

Who publishes the Kev model and under which licence?

Kev is published by Jared Palmer on Hugging Face (jaredpalmer/kev-27b and smaller sizes). The model-card metadata of all seven sizes lists license: apache-2.0. Checked 29 October 2026.

Which Kev sizes exist?

Hugging Face hosts Kev repositories with 0.5B, 0.6B, 0.8B, 4B, 8B, 9B and 27B parameters. According to its model card, Kev 27B is a full-weight fine-tune of Qwen3.8-27B; the smaller sizes are LoRA adapters with a pointer head on Qwen base models.

How does Kev compare with Jev on JevBench?

The Kev table on this page lists dated Capability scores for Kev 27B and other Kev sizes next to TypeSafe Jev 1.13.0 and 1.14.0, from the JevBench data release v1.7.21 (published 29 October 2026). They are different runs from early to mid October 2026, not a ranking and not a same-workload comparison. JevBench is run by the company behind Decision Models.

Can I use Kev on Decision Models?

The /models/kev-27b page offers dedicated hosting of the published Kev 27B weights. A hosted API, a local installer and pre-installed hardware are request forms there. We do not offer Jev.

If you prefer not to operate a model yourself, you can try our existing Jev-compatible models. They use their own model IDs and limits; check answers on your own data.