Biplov GautamTypeSafe AI released Jev on 15 September 2026: the first System One model, an AI decision model that reads a state and answers typed questions (pick an option, score on a scale, or give the probability that something is true) with calibrated probabilities and no generated text. Jev is hosted. You call TypeSafe's API, and the weights are not published.
That leaves room for open models, and they arrived fast. By 24 September there were more than twenty open-weights System One models on this registry, from companies, research groups and individuals. People reach for one to keep data on their own hardware, to cut per-call cost at volume, to fine-tune on their own labels, or simply to know exactly what they are running.
This is a guide to all of them. The live version, updated as models are added, is at Open-source Jev alternatives.
The short answer
"Speaks Jev's API" means the model's own documentation says it serves
TypeSafe's POST /v1/systemone request format, so a client written for Jev
can switch by changing its base URL.
| Model | Maker | Size | Licence | Speaks Jev's API | Worth knowing |
|---|---|---|---|---|---|
| Laya | Convai Innovations | 421M | Apache-2.0 | Yes (laya-serve) | Small encoder, English and multilingual; fine-tunes locally in Laya Studio |
| Kev | Jared Palmer | 0.8B to 27B | Apache-2.0 | Yes | A family of sizes on Qwen3.5 |
| Decider | Mark Marosi | 0.8B to 35B | Apache-2.0 | Yes | The lowest median latency reported, 3.2 ms |
| Bespoke-Nimble-9B | Bespoke Labs | 9B | Apache-2.0 | Yes | Data, recipe and benchmark suite all public |
| Von 1.2 | Victor Hugo Panisa | 395M | Apache-2.0 | Yes | Small encoder; option order doesn't change the answer |
| CLM | Contrastive-LM (Stanford) | 8B | Apache-2.0 | Yes | Also ranks, with a /v1/rank endpoint |
| JevK5 | Alibi Serikbay | 4B | Apache-2.0 | Yes | Teacher labels came from an OpenAI model |
| OpenThai-SystemOne | iApp Technology | 0.8B | Apache-2.0 | Yes | Thai and English, up to 255 options |
| Decision-1.0-Lux-9B | vLLM Semantic Router | 9B | Apache-2.0 | Yes | Runs in vLLM Semantic Router's Decision runtime |
| lev | Interfaze AI | 4B | Apache-2.0 | Yes | A LoRA on Qwen3.5-4B |
Laya also has MLX ports for Apple silicon, laya-mlx and laya-multilingual-mlx.
Why there is no leaderboard here
Every publisher reports accuracy on its own test set: its own held-out split, a JevBench tier, a benchmark built for the release. Bespoke-Nimble-9B reports 90.1% on its 324-example held-out set; Decider reports 80.2% on its 67-task regression set; Von reports 63.9% on the JevBench public standard tier. Those numbers are not comparable, and ranking them would mislead you. Each model's page shows its number next to the suite it was measured on, and that is as far as an honest comparison goes today.
How to choose
1. Licence first. If you are shipping a product, keep to the Apache-2.0 and MIT models. Hopper is research-only, the pngwn scorer is non-commercial, and the licence for Tev1's weights is still being finalised. JevK5 is Apache-2.0, but most of its training labels came from an OpenAI model, so read OpenAI's terms if that matters to you.
2. Already calling Jev? Start with the eleven that speak its API. Point your client at the model's server, run the same requests, and compare the answers with Jev's on your own traffic before switching.
3. Match it to your hardware. The encoder models (Julia 1 at 144M, GLiNER2.5-Decide at 340M, Von at 395M, Laya at 421M, open-jev at 434M) run on a CPU or a small GPU. The ones built on 4B to 12B language models (Kev, JevK5, lev, Nimble, CLM, Winnow) want a GPU, or a Mac with enough memory.
4. Check the languages. Most are English-first. Laya has a multilingual checkpoint, Julia 1 is built on a multilingual encoder, and OpenThai-SystemOne is trained for Thai.
5. Measure on your own labels. Two hundred labelled examples from your own traffic tell you more than any published number. Compare accuracy, and check calibration too: if the model says 0.9, is it right nine times in ten? That is what lets you act on confident answers and send the rest to a person or a bigger model.
Running one
Every model above downloads the same way. Each file is checked against its SHA-256 and cached, so a second pull is instant:
pip install systemonemodels
systemone pull convai-innovations/laya
Each model's page says how to serve it: laya-serve, von-sdk, jevk5-serve, a llama.cpp build, or the maker's own package. Every page also has a playground, so you can try your own questions in the browser before downloading anything.
Fine-tuning one
The main reason to choose an open model is that you can train it on your own decisions. Laya Studio, our Apache-2.0 app, fine-tunes Laya on your own data, on a Mac, Windows or Linux machine. It offers LoRA, DoRA, rsLoRA and LoRA+, scores the fine-tune against the base on held-out rows, and publishes it to the registry in one step:
systemone run studio
We measured what that looks like on a single M4 in MLX vs PyTorch for fine-tuning Laya.
The live list
Models arrive every few days. Open-source Jev alternatives is the same list, generated from the registry, with each model's reported numbers and a head-to-head page against Jev. Every System One model has the hosted ones too.