Jev on Hugging Face
If you're searching Hugging Face for Jev, the honest answer is that you won't find it: Jev is TypeSafe AI's closed-weight, hosted System One model — you call it as an API, and there's no checkpoint, GGUF, or safetensors to pull. What you can find on Hugging Face are the open, self-hostable decision models Jev kicked off — laya, CLM-8B, pplx-decider and others. These are community projects, not Jev and not made by TypeSafe. Here's the real map, with the actual model cards.
Two different things get searched as "jev huggingface": can I download the Jev model from the Hub, and is there any Jev-like decision model on the Hub I can run. The first is no; the second is yes, with several good options. This page covers both honestly and links the real model cards.
Is Jev on Hugging Face?
No. Jev is proprietary and hosted — TypeSafe AI serves it behind an API (you send state plus typed questions and get back calibrated, typed answers), and it hasn't published the weights, training data, or a Hub repo. So there is no official Jev model on Hugging Face, no Jev GGUF to load into text-generation-webui or llama.cpp, and no safetensors to fine-tune. Any repo named "jev" on the Hub is a third-party project, not the hosted model.
That's deliberate. What makes Jev hard to reproduce isn't only the architecture — a non-autoregressive parallel sampler that emits a structured value in one forward pass — it's the RLCD calibration (Reinforcement Learning for Calibrated Decisions) that makes its confidence numbers trustworthy. A downloaded reimplementation can copy the contract and even the architecture, but it doesn't inherit that calibration.
Decision models that are on Hugging Face
Jev defined the category and the open-weight side filled in fast. These all take the same shape — a bounded answer space in, a typed decision out — and ship real Hugging Face model cards. They are separate projects, not Jev and not from TypeSafe, and none are RLCD-calibrated, so you validate confidence on your own data:
- convaiinnovations/laya — Convai Innovations' open System One model (Apache-2.0, 421M) — the most production-oriented open option; weights on the Hub.
- aac6fef/laya-mlx — An MLX build of laya for Apple Silicon — runs on the Mac GPU.
- Contrastive-LM/CLM-v0.1-8B — Open candidate-ranking decision model (Apache-2.0): a frozen Qwen3-8B plus projection heads trained with InfoNCE.
- perplexity-ai/pplx-decider-v1-27b — Perplexity's open-weight decider — a Qwen3-27B fine-tune, Apache-2.0, with a vendor-reported benchmark table on the card.
- Maincode/matilda-jev-v1.5 — A third-party checkpoint named after Jev's contract — community, not TypeSafe; check the card's licence yourself.
How to run an open one from the Hub
There's no Jev repo to clone, but you can pull an open decision model from Hugging Face and serve it behind a TypeSafe-compatible endpoint with Ollaya (the community 'Ollama for decision models'), so code written for Jev points at it by changing one base-URL env var:
# Pull the open weights from Hugging Face
huggingface-cli download convaiinnovations/laya
# ...or let Ollaya fetch + serve it behind a TypeSafe-compatible API
curl -fsSL https://ollaya.dev/install.sh | sh
ollaya run laya --preset triage
# Your Jev client now works by pointing its base URL at the local daemon.You trade calibration and ops for open weights: these models aren't RLCD-calibrated, so a 0.8 doesn't reliably mean right 80% of the time until you've validated it on your own data. See the local-model and Ollama guides for the full self-host path.
If it's hosted Jev you actually want
If you came to the Hub looking to run Jev and what you really want is its calibrated, zero-ops decisions rather than open weights, you don't need to self-host anything. Our hosted endpoint serves the same choice / score / noul API with a jv_live_ key, billed per input token (output free) at well under a cent per decision — try it free in the browser playground first. Exact per-million numbers live on the pricing page; official upstream is POST https://api.typesafe.ai/v1/systemone.
FAQ
Is Jev available on Hugging Face?
No. Jev is TypeSafe AI's closed-weight, hosted System One model — there's no checkpoint, GGUF, or safetensors on the Hub, and no official Jev repo. Any Hub repo named "jev" is a third-party project, not the hosted model.
Can I download Jev weights from the Hub?
No — there's no public Jev checkpoint. If you want something to download and self-host, use an open decision model such as laya (convaiinnovations/laya, Apache-2.0), CLM-8B, or pplx-decider. They accept the same decision shape but aren't RLCD-calibrated, so you validate confidence yourself.
What Jev-like decision models are on Hugging Face?
Several open ones: convaiinnovations/laya (plus an MLX build at aac6fef/laya-mlx), Contrastive-LM/CLM-v0.1-8B, perplexity-ai/pplx-decider-v1-27b, and the community Maincode/matilda-jev checkpoints. All are third-party, not made by TypeSafe, and none are RLCD-calibrated.
How do I run an open decision model from Hugging Face?
Pull the weights with huggingface-cli download, or let Ollaya (the community 'Ollama for decision models') fetch and serve it behind a TypeSafe-compatible endpoint — then your Jev client works by changing one base-URL env var. On a Mac, the laya-mlx build runs on the Apple GPU.
Why isn't Jev on Hugging Face?
TypeSafe keeps Jev closed because the hard, valuable part is the RLCD calibration that makes its confidence trustworthy, not just the parallel-sampling architecture. A downloaded reimplementation can copy the contract but doesn't inherit that calibration, so Jev ships only as a hosted, calibrated API.
See also: Is Jev open source? · Run a local decision model · Jev alternatives · Laya vs Jev · CLM-8B vs Jev · Pricing · Get a key
Want the calibrated model, not a download?
Jev isn't on Hugging Face — it's a hosted, RLCD-calibrated API. Run a real typed decision in the browser with no signup, then grab a jv_live_ key. Sub-cent per decision, output free.