Jev vs Kev 27B
Apache-2.0 LoRA and pointer head on Qwen3.8-27B.
Jev 1.13.0
TypeSafe · hosted System One model
77.1Capability · unranked reference
Intelligence 63.6 · calibration 90.6.
Kev 27B
jaredpalmer · Apache-2.0 LoRA and pointer head on Qwen3.8-27B
78.9Measured · outside the cost cap
Capability 78.9. Outside the ranking cap because estimated cost is $0.529 per 1,000 decisions.
Key differences
- Where it runs JevHosted APIKev 27BSelf-hosted, about 55 GB, bf16
- Licence JevProprietary, hostedKev 27BApache-2.0
- Capability (v1.6.1) Jev77.1 · referenceKev 27B78.9 · measured, not ranked
- Intelligence / calibration Jev63.6 / 90.6Kev 27B64.2 / 93.7
- Contract JevPOST /v1/systemoneKev 27BSame public /v1/systemone contract, local base URL
- Estimated cost Jev$0.032 per 1,000 decisionsKev 27B$0.529 per 1,000 decisions
JevBench v1.6.1, same method
Capability is the mean of intelligence and calibration, ranked only inside Benchmark Heaven’s cost and median-latency caps. Jev 1.13.0 stays on the board as the reference row and is not given a rank. A model can score above that reference and still be left off the numbered list when its estimated cost is too high. Full table ↗
Capability axes, higher is better
Cost $0.529 per 1,000 decisions (estimated). Median latency 0.5 s. Jev’s reference row is $0.032 per 1,000 decisions (not published as a list price), median 0.2 s. Both dollar figures are the benchmark’s, except Jev’s, which Benchmark Heaven marks as not a published list price.
What Kev 27B is
Kev 27B is an Apache-2.0 decision model: a LoRA plus a pointer head on Qwen3.8-27B. One document and a set of typed questions go in; a probability distribution per question comes out in a single forward pass. The author serves TypeSafe’s public /v1/systemone contract, so a TypeSafe-compatible client can point at a local base URL.
On JevBench v1.6.1 it scores Capability 78.9 against Jev’s 77.1. Intelligence is close (64.2 vs 63.6). Calibration is the clearer win (93.7 vs 90.6). The row is unranked because estimated cost is $0.529 per 1,000 decisions — an order of magnitude above Jev’s $0.032 — and median latency is 0.5 s against 0.2 s.
Read this before you switch
- The interesting score sits outside the official ranking because of cost.
- The model card’s serving note is about 55 GB in bf16 on CUDA. Budget a large GPU, not a laptop.
- Smaller Kev checkpoints exist (8B, 4B, 0.6B, 0.5B) and score far lower on the same board. This page is the 27B row.
When to use which
Choose Jev if
- The bill matters. $0.032 vs $0.529 per 1,000 decisions is the gap, not the 1.8 Capability points.
- You want the call hosted, at 0.2 s median, with no 55 GB weight download.
- Intelligence is close enough (63.6 vs 64.2) that hosting is the real difference.
Choose Kev 27B if
- You want the higher calibration number (93.7) and can afford the measured cost.
- You need Apache-2.0 weights that already speak /v1/systemone.
- The deployment is a single large GPU you already operate.
FAQ
Is Kev 27B better than Jev?
On the v1.6.1 Capability score, slightly: 78.9 vs 77.1. Calibration is 93.7 vs 90.6. Intelligence is 64.2 vs 63.6. Estimated cost is $0.529 vs $0.032 per 1,000 decisions, so the board does not rank it.
Does Kev use the Jev API?
It serves the same public /v1/systemone request shape. It is not a TypeSafe product. Point a TypeSafe-compatible client at the local server and set the model name the card documents.
Which Kev is this?
The 27B checkpoint at jaredpalmer/kev-27b, called Kev 1.0 on the JevBench row. The 8B, 4B, and sub-1B Kev rows on the same board are different systems with lower scores.