Jev
Jev Overview
Jev is the first public System One Model from TypeSafe AI (San Francisco) — a commercial, API-only decision model built for automation rather than chat. You send structured questions and get typed decisions with calibrated probabilities and confidence that software can act on directly: set a threshold, act autonomously below/above it, or escalate to a human. It is trained with RLCD (Reinforcement Learning for Calibrated Decisions), a new architecture, sampler, and training algorithm. Jev never generates text, so hallucination cannot occur by construction.
Key Claims (official)
- Decisions, not strings: type-safe typed outputs instead of prose meant for humans.
- Calibrated confidence: every decision includes a confidence estimate for threshold-based automation.
- Speed & cost: 193.6x faster and 444.6x cheaper than LLMs on System One workflows ($0.000081 / 0.114s per decision vs $0.013880 / 8.566s).
- Pricing: $42 per billion input tokens — advertised as 238x lower than Claude Fable 5.1.
- Zero hallucinations: no text generation, so nothing to hallucinate.
Jev vs Laya (open-source alternative)
| Dimension | Jev | Laya |
|---|---|---|
| Availability | Closed commercial API | Open weights, Apache 2.0, self-host |
| Params / backbone | Undisclosed | 421M ModernBERT-large / 322M mmBERT-base |
| Languages | Not specified publicly | 100+ (multilingual checkpoint) |
| Latency / cost | 0.114 s / $0.000081 per decision; $42 per B input tokens | ~33 ms, free when self-hosted |
| typed-decisions | 0.727 published | 0.766 fine-tuned (base zero-shot ~0.36) |
| Banking77 (77 options) | 0.870 | 0.425 at default settings (raise head_max_len) |
| Calibration | Calibrated confidence marketed natively | Ships over-confident (ECE 0.466); refit temperature → 0.081 |
Ecosystem
160+ Jev-style derivatives appeared on Hugging Face within days of launch (open-jev-deberta-v3-large by Kotoba Labs, Open-Jev-9B, Jev-Style-Qwen3.5-2B, JEV-CPU, mini-Jev, and more), plus community leaderboards (jev-decision-index, typed-decision-leaderboard). Laya provides a Jev-compatible HTTP server (pip install "laya[serve]") so Jev-API workflows can migrate without redesign.
Honest Limits
- Weights and architecture are not public — no self-hosting, and official benchmarks use private workflows (agreement with frontier models), so they are not directly comparable to public gold-label scores.
- Lock-in: subject to API pricing and availability policy changes.
Links
Website: https://typesafe.ai/ · Blog: Introducing System One Models & Jev · HF org: https://huggingface.co/TypeSafeAI