Open Source Jev Alternatives: Five Self-Hosted Decision Models to Run Yourself

Description and use

Five open‑source alternatives to Jev offer fast, self‑hosted decision models with low latency, though they trail in accuracy on large option sets.

Problem

Organizations need low‑latency, self‑hosted decision models that match the Jev interface but without paying for frontier models.

Outcome

The projects collectively offer lower latency (e.g., Laya 32.8‑39.5 ms on T4, Kev 47 ms on M3 Pro) and cheaper inference than Jev, though they lag in out‑of‑the‑box accuracy on the 49‑task benchmark (best open entrant 0.704 vs Jev 0.966).

Workflow

• 1. Identify the Jev contract and required answer types (choice, score, noul).
• 2. Choose an open‑source project that matches the desired model size and license.
• 3. Install via pip or clone the repo (e.g., pip install laya).
• 4. Load the model checkpoint (e.g., laya.load("convaiinnovations/laya")).
• 5. Prepare state and question payloads in the Jev format.
• 6. Call predict() or run the provided FastAPI server (POST /v1/systemone).
• 7. Consume the returned answers (choice, score, noul).
• 8. Optionally fine‑tune on a custom dataset or use LoRA adapters for better accuracy.

Evidence

TypeSafe shipped Jev on September 15, 2026 to do exactly that job and nothing else, then kept it closed, hosted, and behind a waitlist.

Source

Laya is the flagship by a wide margin, and the only one of the five you install with a single pip install laya.

Source

Published speed is 39.5 ms per question on a Tesla T4, or 7.2 ms with ten batched.

Source

The largest has 19,301 GitHub stars, and it added over 5,000 of them in the last day alone.

Source

Caveat: on the independent 49-task classifier benchmark, Jev scores 0.966 macro accuracy against 0.704 for the best open entrant.

Source