Open-Jev lands within a point of closed Jev — MIT and self-hosted
The open typed-decision model hits 95.5% on JevBench — within a point of closed Jev — runs local with no tariff, and now reads images too.
Copy markdown
Within a point of closed Jev, and open
On JevBench v1, the openjev-sglang build (Qwen3.6-35B-A3B) scores 95.5%, trailing TypeSafe's closed Jev 1.13 at 96.3% and GPT-5.6 Luna at 97.1% — a gap that sits inside the confidence bands. The project's own Open-Jev-9B reports 97.54% test and 91.97% out-of-distribution accuracy on 26,452 held-out decision rows.
No per-decision tariff
Closed Jev is the cheapest hosted option at $0.027 per 1,000 decisions; the open builds carry no tariff at all when you self-host, at roughly 0.65s median latency — the same tier as the fastest systems in the benchmark. If typed decisions sit anywhere in your eval, grading or moderation loop, that line item can drop to zero.
Typed decisions, three interfaces
Open-Jev returns structured probabilities instead of text through Choice (odds per candidate), Noul (a 0-1 condition scale) and Score (rubric grades), trained on 408,884 typed rows across 25 task sources. It never generates tokens, so the answer is a number you can threshold and the output is free. Build this weekend: swap your LLM-as-judge grader or reranker for it.
It reads images, not just text
The MIT-licensed openjev cross-encoder (Qwen3.5, in 0.8B / 4B / 35B-A3B sizes) scores image claims too, lifting multimodal-claim accuracy from 0.52 to 0.84 — enough for zero-shot visual moderation or reranking without wiring a vision LLM into the loop.
The catch: pin the exact repo
'Open-Jev' is really several efforts under one name — zefan-cai's 2B and 9B (a 27B is still training) and AlexWortega's Qwen3.5 line — built on different bases, and JevBench v1 is a 242-decision pilot. Check the size, base model and license of the specific weights before you route production traffic through them.
Elsewhere: Dream-RSI replays instead of recomputing
Google DeepMind's Dream-RSI lets a discovery agent improve by replaying past attempts against a learned world model instead of re-running expensive searches, cutting agent calls up to 162x with the weights frozen. The code is open. Build this weekend: bolt a replay buffer onto your own search agent.