Open-source alternatives to Jev
Yes, Jev has open alternatives: several open-weight decision models answer the same kind of typed questions on your own hardware, and four take Jev-style questions: Clef, JevK5 and Laya serve its request format, and Kahn1 takes the same question fields on its own route. None has been measured as a drop-in replacement on every task.
Updated October 2, 2026
01 · The short answer
If you want typed decisions with probabilities without sending your data to an API, pick an open decision model. On our 14,663-item held-out set, the most accurate system measured is an open one, Cloudflare's Clef-flash (74.8%), ahead of Jev (73.2%) and Kahn1 4B (72.4%); it needs a GPU with more than 16 GB in bf16. If you want nothing to host, Jev is still ahead of Kahn1 4B on that set. On the 231 public JevBench items, Kahn1 4B, Jev, JevK5 and Clef-flash are level (Kahn1 4B 87.4%, Jev 86.6%, Clef-flash 83.5%, no significant gap). Method and biases: open models.
02 · The open alternatives
| Alternative | Provider | Licence | Sizes | Jev-compatible | Measured here like for like |
|---|---|---|---|---|---|
| Clef, Clef-flash | Cloudflare | Apache 2.0 | 27B, 9B | Yes, announced (Jev API) | Clef-flash: held-out and public JevBench (int8, vs Kahn1) |
| JevK5 v0.3 | allebee | Apache 2.0 | 4B, 9B (+ 2B, Lite) | /v1/systemone request shape | v0.2 on public JevBench (its authors' run and JevBench's) |
| Laya | Convai Innovations | Apache 2.0 | 421M, 322M (encoders) | /v1/systemone request shape (laya-serve) | Held-out and public JevBench (vs Kahn1) |
| Kahn1 | Kahn1 | 4B Apache 2.0, 3B Qwen Research License | 4B, 3B | Jev's question fields under a “schema” key at /v1/evaluate/jev; not a drop-in for Jev clients | Yes: held-out and public JevBench |
| open-alternative-jev (so1) | ikermoel | Apache 2.0 | Any open LLM | No | |
| Tev1-4B-experimental | Together AI | Being finalized | 4B | Held-out and public JevBench (open models) |
The full landscape, hosted and open, is on the landscape page.
03 · What is measured
| Items | Jev 1.13.0 | JevK5 v0.2, JevBench's run | JevK5 v0.2, its authors' run | Clef-flash (int8) | Laya | Kahn1 4B | |
|---|---|---|---|---|---|---|---|
| Kahn1 held-out, all primitives (Choice over the same 8 options) | 14,663 | 73.2% | 74.8% | 59.1% | 72.4% | ||
| Public JevBench, all tiers | 231 | 86.6% | 85.3% | 86.1% | 83.5% | 57.6% | 87.4% |
| Public JevBench, hard tier | 111 | 73.0% | 73.9% | 67.6% | 32.4% | 75.7% |
Same items, same options, same labels, but three separate runs: Jev's outcomes are the ones JevBench publishes, JevK5's come from its authors' published run and from JevBench's own run, and Kahn1's from ours (k = 3 option orders, calibrated). Paired with JevK5's own published run, Kahn1 4B gets 202 of 231 items right and JevK5 199 (13 against 10 items only one gets right, exact McNemar test, p = 0.68); against Jev's published outcomes (200 of 231), 13 against 11, p = 0.84; in JevBench's own run, JevK5 v0.2 gets 197 (85.3%). Details: benchmarks.
04 · Switching from Jev
The three question types carry over as they are: choice, score and noul. Clef, JevK5 and Laya serve Jev's request format, POST /v1/systemone. Kahn1 takes the same question fields (type, instructions, criteria) under a schema key at POST /v1/evaluate/jev, while a Jev client sends questions and model to /v1/systemone: plan a small adapter. Recalibrate on a few hundred of your own examples before reusing thresholds tuned for Jev.
$ SYSONE_MODEL=Okura66/Kahn1-Qwen3.5-4B uv run sysone serve --port 8000 $ curl -X POST http://127.0.0.1:8000/v1/evaluate/jev -H "Content-Type: application/json" \ -d '{"state": "Hello, I cannot log in to my account.", "schema": {"category": {"type": "choice", "instructions": "Support ticket category", "criteria": {"bug": "Something is broken", "account": "Login, access"}}}}'
05 · When Jev is still the right pick
- You want overall accuracy close to the best measured on our held-out set, with no infrastructure to run.
- Your texts are long: Jev takes 64k tokens per request (32k for the state plus the longest question), per its documentation.
- You work in English, where TypeSafe says Jev is most accurate.
- You are fine with a third party processing your data and a price per input token.
06 · FAQ
Is Jev open source?
No. Jev is a hosted API from TypeSafe; its weights are not published. You call it at api.typesafe.ai and pay per input token ($0.042 per million, output free, per its documentation).
Can I run Jev locally?
Not Jev itself. Open decision models run on your own hardware: Clef and Clef-flash (Cloudflare), JevK5, Laya, Kahn1 and others, several of them under Apache 2.0.
Which open alternative is closest to Jev?
It depends on the task, and no open model has been measured as a drop-in replacement on every task. In JevBench's own runs on the 231 public items (v1.4), several systems with public code or weights score at or above Jev's 86.6%: NInfer Qwen3.8-Flash-Next and JevOne (89.6%), OpenJev (thinking) and swanOne (88.7%), djev (thinking, 87.4%), reflex-27b (87.0%) and SimpleJev Qwen3.8-27B (86.6%); JevK5 v0.2 scores 85.3%. Kahn1 4B scores 87.4% on the same items in our own run, which JevBench has not reproduced; paired against Jev's published outcomes the gap is not significant (exact McNemar p = 0.84). On Kahn1's 14,663 held-out items, Clef-flash (run in int8 on 16 GB) scores 74.8%, Jev 1.13.0 73.2%, Kahn1 4B 72.4%, Tev1 72.0% and Laya 59.1%; on JevBench, Clef-flash 83.5%, Tev1 76.2% and Laya 57.6%.
Which alternatives accept Jev's request format?
Clef is announced as fully Jev-API compatible, and JevK5 and Laya (through laya-serve) serve the POST /v1/systemone request shape. Kahn1 takes the same question fields (type, instructions, criteria) under a "schema" key at POST /v1/evaluate/jev, so a Jev client needs a small adapter.
Is Kahn1 affiliated with TypeSafe?
No. Kahn1 is an independent project; it takes Jev's question fields and compares itself with Jev on the same items.