Jev alternatives, ranked and explained
Jev is TypeSafe's hosted System One model: state plus typed questions in, calibrated structured decisions out. Every serious alternative on the independent JevBench v1.4.2.2 leaderboard is an open-weight rebuild you run yourself — each trades Jev's calibration, option limits and zero-ops hosting for your own GPU. Here is the whole field, honestly ranked.
- 95systems on JevBench v1.4.2.2 · 91 ranked
- 3open 4B models outrank Jev on the composite
- 53.1Jev's intelligence score — still #1 in the top ten
- 255 / 64kJev's option cap and context — most alternatives cap far lower
The field, by JevBench rank
Scores are the JevBench v1.4.2.2 composite — a harmonic mean of intelligence, calibration, speed and cost, measured on 842 decisions per system. Every page links the project's own repository; nothing here is copied from vendor marketing.
How to choose
- Hosted vs self-hosted — Jev is the only generally-available hosted decision model at the top of the board. Every alternative means GPUs, quantisation and version pinning on your side.
- Accuracy where it hurts — read the hard and sealed tiers, not the composite. Jev's 74.1% hard / 36.7% sealed still lead everything below #3.
- Calibration — if you threshold on probabilities, prefer a fitted model (Jev 76.3, Imajev 80.4) over normalised logits (Winnow 64.8, SemIf 66.8).
- Interface limits — Jev takes up to 255 options and 64k tokens. Alternatives cap between 16 and 26 options and 8–32k tokens.
- Licence — Apache-2.0 (Imajev, decider, JevK5), MIT over open weights (SemIf), Gemma terms (Cygnet, Winnow). Check before commercial deployment.
- Multimodal — only Imajev-4B and Winnow read images; Jev is text-only.
The honest summary
Three open 4B models now outrank Jev on the official composite — real, measured results, not marketing. They win it on speed and cost axes that a networked API cannot match. On the accuracy views the order flips: Jev is #3 when accuracy is weighted 60% and #1 on intelligence alone, with the best sealed-set robustness below the top two.
Pick an alternative when your constraint is deployment — data that cannot leave your network, a hard cost ceiling, images in the request, or Apache/MIT terms. Pick Jev when your constraint is the decision itself: ambiguous cases, calibrated probabilities you can threshold on, wide option sets, and a versioned API with nothing to operate.
FAQ
What is the best Jev alternative?
On the JevBench v1.4.2.2 composite, Imajev-4B (#1, 67.4) is the highest-ranked alternative — an Apache-2.0 Qwen3.5-4B LoRA that also reads images. Plumb-4B (#2, 65.8) and decider-4b (#3, 64.1) follow. All three are self-hosted; none beats Jev on the intelligence-only view.
Are Jev alternatives free?
The open-weight ones are free to download — Imajev-4B, decider-4b and JevK5 are Apache-2.0, SemIf is MIT over open weights. You pay in GPU time instead of API calls. Hosted alternatives like djev bill per use, same model as Jev.
Can I self-host Jev itself?
No — TypeSafe does not publish Jev weights. If on-prem inference or air-gapped deployment is a hard requirement, you need one of the open-weight alternatives below. If a managed API is acceptable, Jev is the hosted option in this category.
Which alternative is easiest to start with?
decider-4b (~8.4 GB bf16) and JevK5 (~9 GB) fit a single consumer GPU and speak the same state-plus-questions contract. Cygnet needs no training at all but a bigger GPU. SemIf is the most readable codebase — MIT over stock Qwen3.5-4B logits.