Loading the alternatives…
OPEN ALTERNATIVES TO JEV
Jev, from TypeSafe AI, is a hosted System One model: an AI decision model that returns typed answers with calibrated probabilities instead of text. Its weights are not published. These are the models you can download, run on your own hardware and fine-tune instead: 30 from 25 makers, listed from the live registry.
For each: what it decides, its size and licence, the benchmark its publisher reports with the suite named, and a head-to-head page against Jev. Updated November 20, 2026.
These serve TypeSafe’s POST /v1/systemone request format, according to their own documentation, so a client written for Jev can switch by changing its base URL.
| Model | Decides | Parameters | Licence | Reported accuracy | Median latency | Released | Against Jev |
|---|---|---|---|---|---|---|---|
| layaConvai Innovations | choice, score, noul, classify, route | 421M | apache-2.0 | — | 39.5 ms | Sep 18, 2026 | — |
| deciderMapika | choice, score, noul, classify, route | 1.9B | apache-2.0 | 75.2%Decider regression set, 28 held-out tasks (decider-2b v11) | — | Sep 24, 2026 | — |
| clmContrastive-LM | choice, score, noul, rank, classify, route | 8.0B | apache-2.0 | — | — | Sep 23, 2026 | — |
| startlux-decisionStartLux | choice, score, noul | 4.7B | cc-by-nc-4.0 | 88.3%JevBench public set (231 items; 204 correct), maker's run | 26 ms | Sep 29, 2026 | — |
| lumma-fevFrontiersMind | choice, score, noul | 649M | apache-2.0 | 64.0%typed-decisions (maker's table; split not stated) | 45.8 ms | Sep 22, 2026 | — |
| kevJared Palmer | choice, score, noul, classify, route | 4.0B | apache-2.0 | 83.8%transfer-v4 (locked, out of domain) | — | Sep 19, 2026 | — |
The same kind of model, with their own interface: a Python package, a different endpoint, or a chat-completions wrapper. Each model’s page shows how to call it.
| Model | Decides | Parameters | Licence | Reported accuracy | Median latency | Released | Against Jev |
|---|---|---|---|---|---|---|---|
| julia-1Supersonic Labs | choice, score, noul, classify, route | 144M | apache-2.0 | 73.2%typed-decisions test set (400 cases, 2,000 questions) | — | Sep 23, 2026 | — |
| vegaNandakishor M | choice, score, noul, classify, route | 800M | apache-2.0 | 55.6%His 11 October 2026 benchmark, published with its results and its harness (a notebook in vegaml): macro accuracy over 43 public tasks for the 0.8B, 0.717 for the 4B; with JevBench (231 items) and DecisionBench (1,075 rows), 3,841 decisions in all, seed 20261011. On his own set of 2,050 choice decisions over six intent workflows (esci, clinc150, appliances, banking77, when2call, apibank), one choice question each, which he has not published with its question wording, he reports 0.763 (calibration error 0.026) for the 0.8B with its public_intents adapter, 0.389 without the adapter and 0.803 for the 4B with it. He publishes no accuracy for the typed_decisions adapters | — | Oct 9, 2026 | — |
| miraSAGEA | choice, score, noul, classify, route | — | apache-2.0 | 61.3%s1-decision-bench | 30 ms | — | — |
| caliber-flex-5bonepass | choice, noul, score | — | other | — | — | — | — |
| caliber-flex-1bonepass | choice, noul, score | — | other | — | — | — | — |
| laya-typed-decisions-noulxpConvai Innovations | choice, score, noul, classify, route | 421M | apache-2.0 | — | 126 ms | — | — |
| laya-noulxp-cpuConvai Innovations | choice, score, noul, classify, route | 421M | apache-2.0 | — | 116 ms | — | — |
| gliner2-5-decideFastino Labs | choice, score, noul, classify, extract, route | 340M | apache-2.0 | 60.2%Fastino fast-decisions suite (17 datasets) | 38.3 ms | Sep 24, 2026 | — |
| laya-noulxp-gpuConvai Innovations | choice, score, noul, classify, route | 421M | apache-2.0 | — | — | — | — |
| laya-multilingual-noulxpConvai Innovations | choice, score, noul, classify, route | 322M | apache-2.0 | — | 44.9 ms | — | — |
| neohorse-jevTokenRhythm | choice, score, noul, classify, route | 4.0B | apache-2.0 | 75.3%JevBench public set (231 items), vLLM, maker's run | — | Sep 23, 2026 | — |
| rizzo-flowRizzo AI Academy | choice, score, noul, classify, route | 4.0B | apache-2.0 | 64.8%LocalLLaMA/typed-decisions test split (400 cases, 2,000 decisions), Q8_0 | 195 ms | Sep 25, 2026 | — |
| open-jev-deberta-v3-largeKotoba Labs | choice, score, noul, classify | 434M | apache-2.0 | 85.4%Kotoba held-out test, seen question types (banking77, SST-5, BoolQ) | 28 ms | Sep 18, 2026 | — |
| waldORG2 AI | choice, score, noul, classify, route | 4.2B | apache-2.0 | 88.3%JevBench public set (231 items; 204 correct), maker's run with the JevBench CLI, v1.2, effort none; items used as a development scoreboard | 33 ms | Oct 1, 2026 | — |
| anyjevNokia | choice, score, noul, classify, route | 8.0B | apache-2.0 | 77.1%LocalLLaMA/typed-decisions (2,000 held-out decisions), L2 with 300 labels per question | — | Sep 21, 2026 | — |
| canopy-jevRuoxi Qiu | choice, noul, score | 27B | apache-2.0 | 87.4%JevBench public set (231 items; 202 correct), the maker's own fresh zero-shot run with the upstream scorer | — | Oct 2, 2026 | — |
| biodecisionDimitris Papadopoulos | choice, score, noul, classify | 4.0B | other | 75.3%BioDecision v2 held-out test set (63,432 overlap-screened decisions from public biomedical benchmarks; the maker's own split) | 39 ms | Sep 28, 2026 | — |
| nativ-he-decisionYoav Pinto | choice, classify, route | 364M | cc-by-4.0 | 96.2%nativ-bench user-intent task, 4 options (MASSIVE Hebrew, 2,973 items; the maker's own suite) | — | Oct 4, 2026 | — |
| docto-decisionBofeng Huang | choice, score, noul | 4.7B | apache-2.0 | 88.9%Docto Decision Bench fr v0.1 (12 tasks, mostly silver labels; the maker's own benchmark) | 47 ms | Oct 2, 2026 | — |
| gutsyOmar (kouhxp) | choice, score, noul, classify, route | 800M | apache-2.0 | 73.2%JevBench public set (231 items; 169 correct), the maker's own run with Q8_0 | — | Oct 2, 2026 | — |
| bekko-system-oneYuichi Tateno (hotchpotch) | choice, noul, score | 395M | other | — | 5 ms | Sep 29, 2026 | — |
| bev-deciderAvishek Biswas | choice, score, noul | 478M | cc-by-nc-4.0 | 74.7%avbiswas/bev-decision test split (5,000 held-out questions over 2,617 states), the maker's own | — | Sep 28, 2026 | — |
| modernjev-decideMaziyar Panahi | choice, rank, route | 150M | apache-2.0 | 73.4%AgentToolDecisions-180K held-out test, next action type (1,158 decisions, 3 choices; the maker's split) | — | Sep 30, 2026 | — |
| ajevAndy Zhang | choice, score, noul | 26B | apache-2.0 | — | 49 ms | Oct 5, 2026 | — |
Accuracy and latency are what each publisher reports, on its own suite and hardware, so they are not a ranking. The reliable comparison is your own labelled examples. Every System One model, hosted ones included, is in one table.
Download any of them from your terminal. Every file is checked against its SHA-256 and cached.
pip install systemonemodels systemone pull convai-innovations/laya
System One Studio fine-tunes Laya on your own data, on a Mac, Windows or Linux machine, and proves the result beats the base before you publish it.
systemone run studio
No. Jev is TypeSafe AI’s hosted System One model: you call it over an API, and its weights are not published. The 30 models on this page are the open alternatives: you download the weights and run them yourself.
There is no fair ranking yet: each publisher reports accuracy on its own suite, so the numbers are not comparable. Choose by your constraints instead. If you already call Jev, start with the ones that speak its API: laya, decider, clm and startlux-decision. The smallest are julia-1 (144M), modernjev-decide (150M) and laya-multilingual-noulxp (322M). The fastest reported are bekko-system-one (5 ms), startlux-decision (26 ms) and open-jev-deberta-v3-large (28 ms), on their publishers’ own hardware. Then measure the shortlist on your own labelled examples.
laya, decider, clm, startlux-decision, lumma-fev and kev. Their own documentation says they serve TypeSafe’s POST /v1/systemone request format, so a client written for Jev can switch by changing the base URL. Check each model’s page for its server and the question types it supports.
Open weights under a licence that allows commercial use: julia-1 (apache-2.0), laya (apache-2.0), vega (apache-2.0), mira (apache-2.0), laya-typed-decisions-noulxp (apache-2.0), laya-noulxp-cpu (apache-2.0), gliner2-5-decide (apache-2.0), laya-noulxp-gpu (apache-2.0), decider (apache-2.0), laya-multilingual-noulxp (apache-2.0), clm (apache-2.0), neohorse-jev (apache-2.0), rizzo-flow (apache-2.0), open-jev-deberta-v3-large (apache-2.0), wald (apache-2.0), lumma-fev (apache-2.0), kev (apache-2.0), anyjev (apache-2.0), canopy-jev (apache-2.0), nativ-he-decision (cc-by-4.0), docto-decision (apache-2.0), gutsy (apache-2.0), modernjev-decide (apache-2.0) and ajev (apache-2.0). Others publish weights under a non-commercial or research-only licence; each model’s page names its licence.
Yes. pip install systemonemodels, then systemone pull namespace/name downloads any of them, with every file checked against its SHA-256. Laya also has MLX ports for Apple silicon, and System One Studio fine-tunes it on your own Mac, Windows or Linux machine (systemone run studio).
Yes, which is the main reason to choose an open one. System One Studio fine-tunes Laya with LoRA, DoRA, rsLoRA or LoRA+, scores the result against the base on held-out rows, and publishes it to the registry. The other models ship their own training code; systemone push publishes what you train.
Read on: What is an AI decision model? · The open reproductions, in depth · How Jev works · Jev alternative, defined