Self-hosted Jev alternatives
Every model on this page runs on your hardware under Apache 2.0 or MIT. The latency column is per three-question case on a four-core CPU VM with no GPU, which is the floor; a GPU moves most of these under 50 ms.
| # | Model | Decisions | Beats baseline | ECE raw → held-out fitted | p50 / case | Extraction F1 |
|---|---|---|---|---|---|---|
| 1 | GLiNER2 largeFastino | 0.716 | 6/7 | 0.261 | 627 ms† | 0.968 |
| 2 | GLiNER 2.5 baseFastino | 0.643 | 5/7 | 0.251 | 332 ms† | 0.968 |
| 3 | GLiClass large v3.0Knowledgator correct payload, no prompt | 0.642 | 5/7 | 0.250 | 977 ms | n/a |
| 4 | GLiNER2 baseFastino | 0.631 | 3/7 | 0.350 | 190 ms† | 0.935 |
| 5 | GLiClass multilang ultraKnowledgator criteria + single-hypothesis Noul (correct payload) | 0.628 | 4/7 | 0.185 | 672 ms | n/a |
| 6 | GLiClass base v3.0Knowledgator criteria + single-hypothesis Noul (correct payload) | 0.606 | 5/7 | 0.146 | 300 ms | n/a |
| 7 | GLiClass modern large v3.0Knowledgator criteria + single-hypothesis Noul (correct payload) | 0.602 | 4/7 | 0.234 | 572 ms | n/a |
| 8 | GLiNER 2.5 smallFastino | 0.574 | 4/7 | 0.172 | 132 ms† | 0.812 |
| 9 | GLiClass instruct base v1.0Knowledgator criteria + single-hypothesis Noul (correct payload) | 0.559 | 3/7 | 0.390 | 323 ms | n/a |
| 10 | GLiClass instruct large v1.0Knowledgator criteria + single-hypothesis Noul (correct payload) | 0.545 | 3/7 | 0.308 | 1.0 s | n/a |
| 11 | GLiGuard 300MFastino | 0.537 | 3/7 | 0.230 | 155 ms† | 0.000 |
| 12 | LAYAConvai Innovations | 0.520 | 3/7 | 0.064 | 378 ms† | n/a |
| 13 | Vonwfzyx | 0.509 | 4/7 | 0.234→ 0.122 | 597 ms | n/a |
| 14 | GLiNER 2.5 multiFastino | 0.494 | 4/7 | 0.244 | 346 ms† | 0.885 |
| 15 | LAYA typed-decisionsConvai Innovations | 0.475 | 3/7 | 0.131 | 375 ms† | n/a |
| 16 | GLiNER2 multiFastino | 0.464 | 2/7 | 0.354 | 202 ms† | 1.000 |
| 17 | LAYA multilingualConvai Innovations | 0.431 | 2/7 | 0.284 | 149 ms† | n/a |
| 18 | OpenJEVopen-jev | 0.427 | 3/7 | 0.210 | 50 s† | n/a |
| 19 | GLiClass edge v3.0Knowledgator criteria + single-hypothesis Noul (correct payload) | 0.418 | 2/7 | 0.287 | 41 ms | n/a |
| 20 | GLiNER multitask v1.0Knowledgator class names as labels | 0.378 | 1/7 | 0.223 | 1.8 s | n/a |
Ranked by decision accuracy. Metatext decisions: Three hand-written workflows (support tickets, security incidents, agent traces), seven questions, about 123 judgments per model. Small, so treat a ±0.1 difference as a tie; every question has a majority baseline the model has to beat. Every model is zero-shot on this suite. Decisions is the task-weighted mean accuracy of the best configuration per model; ECE is the mean expected calibration error over tasks, and the arrow shows the held-out ECE after one temperature per primitive fitted on the other half. Latency is p50 per three-question customer-service case on four CPU cores (Jev: API round trip); † means taken from the benchmark run rather than a clean timing pass. Updated September 2026.
Same ranking on the other suite: typed-decisions.
Jev-shaped models not yet run
Surveyed on Hugging Face; each takes the System One request. Their own published figures are on their model pages, attributed; none is comparable to the table above until it has a row.
- Decider 2B1.9B · Apache 2.0
A decoder-backbone System One reproduction with a proper-scoring-rule SFT stage and a calibration-aware RL stage; serves the /v1/systemone request shape.
- open-jev DeBERTa-v3-large435M · Apache 2.0
Single-pass encoder with a per-option head over state and all questions in one sequence; 512-token context with the state cut to 256 tokens.
- Gavel base150M · MIT
Von's mechanism trained deliberately for the decision task, small and multilingual, with a fitted temperature and the card's own negative results.
- Kev 0.6B0.6B · Apache 2.0
A reproduction of the architecture described in "Jev's Architecture Unmasked"; the only one that measured the real Jev on its own frozen dev sets and published a research log.
- Jev schema scorer435M · Apache 2.0
Scores the literal Jev question JSON as the second sequence, so criteria are read at inference.
- System One Qwen3.5 4B scorer4.2B · Non-commercial
Pointwise scorer trained on 46k System One questions across nine families; reports its calibration separately from its accuracy.
- Qwen-Jeff B21.7B · Apache 2.0
Pointwise scorer with a CLI and an "automate if p ≥ 0.9588" gate calibrated to at most 5% error.
- RLCD ModernBERT 151M151M · Apache 2.0
A GLiClass-mechanism model with an RLCD-style objective; small enough to run in the browser (ONNX + WebGPU demo).
Frequently asked questions
Which self-hosted model is best?
GLiNER2 large on accuracy (0.716 on the Metatext suite, 0.968 extraction F1); LAYA on calibration (ECE 0.064 untuned, the best of any open model); GLiClass edge on latency (about 40 ms per case at 33M parameters).
How fast are they on CPU?
Between 40 ms (GLiClass edge) and about one second (GLiClass large, GLiNER2 large on long states) per three-question case on four cores. OpenJEV on a 0.5B LLM is the outlier at about 50 seconds.
Related
Every open-weights alternative to TypeSafe's Jev, ranked by measured decision accuracy on the 2,000-decision typed-decisions suite: LAYA, Von, GLiNER2, GLiClass, OpenJEV, plus the Jev-shaped checkpoints not yet run.
System One models ranked by p50 latency per three-question case on a four-core CPU, against Jev's hosted round trip.
System One models ranked by expected calibration error on the Metatext suite, untuned, with the held-out ECE after temperature scaling where it was fitted.