Jev vs Kev
Kev is an Apache-2.0 family of small decision models that speak the same /v1/systemone contract — you host them. Jev is hosted. Same request shape, different ownership of weights and ops. Independent hosted gateway; not the official TypeSafe site.
What Kev is
Kev, by Jared Palmer, fine-tunes Qwen bases into decision models with the same question types. Published sizes run from 0.8B — which runs on an Apple Silicon Mac — up to larger GPU checkpoints, and each checkpoint ships a fitted temperature. The server speaks the System One request shape, so an existing client keeps its payloads and swaps the base URL.
Same request, your machine
The body can look identical: a state, plus choice, score and yes/no questions, and back comes a distribution. What changes is everything around it: the model version you pin and patch, the GPU you keep warm, the calibration you re-check on every upgrade, and the data path that never leaves your estate. That is real control, with a real ops bill.
Cost and scale
Self-hosting Kev is effectively free per call once the hardware is running; it wins on high steady volume, air-gapped requirements, or Mac-only fleets. The hosted path is per-use — 1 credit per 1,000 input tokens, $10 for 100,000 credits, credits that never expire — and hands versioning, capacity and failure behavior to Optio. Your bug list changes shape: instead of serving, you tune criteria and thresholds.
When to pick which
Pick Kev when the request shape on hardware you control is the requirement. Pick Jev through Optio when you want that same request shape against hosted Jev — keys, prepaid credits, no GPU — and you would rather spend engineering time deciding what your criteria say than keeping a decision server alive.
Strengths and honest limits
Jev's honest limits: you cannot patch its weights or run it on your own metal — that is the product constraint, not a technical one. Kev's flip side: pinning one checkpoint is also freezing it. Decide which risk you would rather own — dependency on a hosted endpoint, or dependency on your own serving stack staying calibrated.
Quick comparison
| Question | Jev | Kev |
|---|---|---|
| Weights | Proprietary, via Optio endpoint | Apache-2.0 checkpoints, 0.8B upward |
| Hardware | None | Apple Silicon or your GPUs |
| Operations | None — endpoint, keys, credits | Serving, versioning, calibration checks |
| Cost shape | Per use: 1 credit ≈ 1,000 input tokens | Hardware + your time; near-zero marginal cost |
FAQ
- Can I point my existing Jev client at Kev?
- Often yes — Kev speaks the same /v1/systemone request contract, so the payload mostly survives the swap. The answers will differ, because the weights differ: re-validate on your cases and re-fit thresholds before trusting it.
- Do Kev and Jev weigh in at the same quality?
- Not necessarily — different models, different training. At 0.8B Kev trades quality for hardware cost; judge both on your own ten-case validation set rather than any claims on either site.
- Which is cheaper over a year?
- Kev at high steady volume — amortize its GPU and your ops time. Jev on credits at low or spiky volume, or whenever nobody on your team wants to serve a model.
- Is this the official Jev site?
- No. Optio is an independently operated gateway for calling Jev and is not affiliated with TypeSafe or with any alternative model mentioned on this page.