# Kev as a Jev alternative

> Kev is an open Jev-like model family on Qwen with a System One compatible server. Compare sizes, hardware, accuracy, license and migration from Jev.

Last verified: 2026-10-12.

## Answer

Kev is Jared Palmer's open-weight family trained on Qwen 3.5 and Qwen 3.8 bases. kev.serve exposes POST /v1/systemone so the TypeSafe Python SDK can target localhost. README publishes vendor-reported accuracy and Brier tables against hosted Jev on frozen suites. Independent Decision Index scores Kev-27B slightly below Jev in our snapshot, but self-host economics and fine-tuning freedom drive most evaluations.

Kev ships Apache 2.0 decision checkpoints with a TypeSafe-compatible HTTP server, so many teams only change baseURL while keeping parallel Choice, Score, and Noul questions.

## At a glance

| Topic | Kev | Jev | Source |
| --- | --- | --- | --- |
| Maker | Jared Palmer | TypeSafe AI | [GitHub: jaredpalmer/kev](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi) |
| First public release | Public GitHub releases from 2025 | Hosted System One (ongoing) | [GitHub: jaredpalmer/kev](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi) |
| License / access | Apache 2.0 | Hosted API (weights not downloadable) | [Kev README license](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi#license) |
| Hosted price (input) | not published | $0.042 per M input tokens (vendor list, 2026-10-12) | [TypeSafe models page (jev-1.13.0 price table)](https://docs.typesafe.ai/models.md?ref=jev.aitools.fyi) |
| API shape | POST /v1/systemone via kev.serve | POST /v1/systemone | [Kev README usage; Jev: TypeSafe API reference](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi#usage) |
| Question types | Choice, Score, Noul | Choice, Score, Noul | [Kev README usage](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi#usage) |
| Inputs | Single state string plus questions map | Structured state string | [Kev README usage](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi#usage) |
| Model sizes | 0.8B, 4B, 9B, 27B | Hosted Jev (closed) | [Hugging Face: kev-4b](https://huggingface.co/jaredpalmer/kev-4b?ref=jev.aitools.fyi) |
| Latency | not published | not published | [GitHub: jaredpalmer/kev (no published p95 SLA cited)](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi) |
| Decision Index (public) | 58.78 (Decision Index 0.3.1, snapshot 2026-10-10; independent harness, not vendor-only) | 60.11 (Decision Index 0.3.1, snapshot 2026-10-10; independent harness, not vendor-only) | [Decision Index 0.3.1 public snapshot](https://huggingface.co/spaces/multimodalart/jev-decision-index/resolve/main/data/v03.json?ref=jev.aitools.fyi) |
| Data handling | not published | Processed under TypeSafe terms | [Self-hosted data handling not published on a single vendor page](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi) |
| SDKs | typesafe_sdk pointed at local base URL | typesafe_sdk, gateways, MCP ecosystem | [Kev README usage](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi#usage) |
| Adoption signal | 8,952 GitHub stars (2026-10-12); 21,656 HF downloads (2026-10-12) | not published | [GitHub repository (stars)](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi) |

## What Kev is

Kev is Jared Palmer's open-weight family trained on Qwen 3.5 and Qwen 3.8 bases. kev.serve exposes POST /v1/systemone so the TypeSafe Python SDK can target localhost. README publishes vendor-reported accuracy and Brier tables against hosted Jev on frozen suites. Independent Decision Index scores Kev-27B slightly below Jev in our snapshot, but self-host economics and fine-tuning freedom drive most evaluations.

## Accuracy and benchmarks

Independent Decision Index edition 0.3.1 lists Kev at 58.78 in our imported snapshot.

## Fit and tradeoffs

Choose Kev when your constraints match its deployment model: You must run weights inside your VPC without sending state to TypeSafe; You want fine-tune adapters on your own labels; You already operate GPU inference and can own calibration.

## Pricing and ops

Self-host economics are GPU hours, checkpoint storage, and on-call time; vendor README tables against hosted Jev are directional, not your invoice.

## Migrating from hosted Jev

Keep your question map and threshold logic stable. Swap client configuration to point at the new base URL or SDK, then run shadow traffic until disagreement rates fall below your risk budget.

```typescript
// Before: hosted Jev
const client = new TypeSafe({ apiKey: process.env.TYPESAFE_API_KEY });
await client.systemone({ model: "jev-latest", state, questions });

// After: self-hosted System One compatible server (Kev, decider, SemIf, etc.)
const client = new TypeSafe({ baseURL: "http://localhost:8000" });
await client.systemone({ model: "kev-latest", state, questions });
```

## Ecosystem links

Use the jaredpalmer/kev catalog listing and kev.serve README for SDK wiring; sibling pages on this hub compare other System One servers.

## Product-specific notes

kev.serve is the supported path for POST /v1/systemone compatibility; pick a checkpoint size that matches your GPU budget before you promise 27B accuracy. README tables that compare to hosted Jev are vendor-run on frozen suites; treat them as directional, not as your compliance sign-off. Fine-tuning adapters on your labels is the main reason teams pick Kev over metered tokens, not a one-point Decision Index gap. Kev checkpoints span sub-billion to 27B parameters; match VRAM, batch size, and concurrency before you copy a demo config. Shadow mode against hosted Jev should log raw probabilities, not just argmax labels, so you can diff calibration on disputed rows. Keep rollback to api.typesafe.ai for at least one release when you move a production gate to self-hosted Kev. Vendor README and Spaces demos are fine for quick looks; load-test kev.serve on your own hardware before you promise capacity.

## When to choose Kev

- You must run weights inside your VPC without sending state to TypeSafe
- You want fine-tune adapters on your own labels
- You already operate GPU inference and can own calibration

## When to stay on Jev

- You want zero GPU ops and predictable hosted billing
- You need the vendor-supported api.typesafe.ai path for compliance
- Your option sets exceed what you want to tune on open checkpoints

## In the directory

- [Jev alternatives hub](https://jev.aitools.fyi/alternatives)
- [System One primer](https://jev.aitools.fyi/learn/system-one)
- [What Jev is](https://jev.aitools.fyi/learn/jev-typesafe)
- [Where to run Jev](https://jev.aitools.fyi/learn/where-to-run-jev)
- [Integrations](https://jev.aitools.fyi/categories/integrations)
- [Benchmarks](https://jev.aitools.fyi/categories/benchmarks)
- [Kev](https://jev.aitools.fyi/tools/jaredpalmer-kev)
- [Decision Index](https://jev.aitools.fyi/benchmarks/apolinario-decision-index)
- [Jeeves](https://jev.aitools.fyi/tools/posthog-jeeves)

## FAQ

### Is Kev the same product as TypeSafe Jev?

No. Jev is TypeSafe's hosted System One service. Kev is a separate product with its own weights, terms, and ops model unless explicitly documented as compatible HTTP.

### Does Kev beat Jev on Decision Index?

The public snapshot in our metrics file scores Kev; compare the facts table to hosted Jev. Edition 0.3.1 scores Kev-27B; your fine-tunes and batch sizes can move accuracy away from that headline number.

### Can I use the TypeSafe SDK with Kev?

Point typesafe_sdk at kev.serve with a pinned checkpoint; verify model strings and probability metadata against hosted Jev in staging shadow mode.

### Should I leave hosted Jev entirely?

Not necessarily. Teams often keep Jev for gates that need vendor support while experimenting with Kev on lower-risk paths. Run shadow mode until metrics prove parity.

## Sources

- [GitHub: jaredpalmer/kev](https://github.com/jaredpalmer/kev?ref=jev.aitools.fyi)
- [Hugging Face: jaredpalmer/kev-4b](https://huggingface.co/jaredpalmer/kev-4b?ref=jev.aitools.fyi)
- [Hugging Face Space: jaredpalmer/kev](https://huggingface.co/spaces/jaredpalmer/kev?ref=jev.aitools.fyi)
- [TypeSafe docs: System One / Jev API](https://docs.typesafe.ai/?ref=jev.aitools.fyi)
- [Decision Index public snapshot (edition 0.3.1)](https://huggingface.co/spaces/multimodalart/jev-decision-index/resolve/main/data/v03.json?ref=jev.aitools.fyi)

Canonical: https://jev.aitools.fyi/alternatives/kev
Markdown: https://jev.aitools.fyi/alternatives/kev.md