OpenJev v0.2.0

OpenJev v0.2.0: the open-source, local alternative to Jev. This release updates the 4B model.

Highlights

	v0.1	v0.2
Live-website tasks completed (153 held-out tasks)	30 / 153	39–41 / 153
Browser steps on unseen websites (Mind2Web step success)	75.9	78.8
Out-of-domain general decisions (Kev transfer-v4)	73.8	81.4
wev-bench step success (2,346 frozen browser steps)	68.0	73.1
Calibration error (ECE)	0.02–0.06	0.01–0.04
Browser step latency (RTX 5090, p50)	210 ms	184 ms

• Model: alanhuangya/OpenJev-4B. It runs on a Qwen3.5-4B backbone and serves calibrated probabilities. v0.1 stays available as revision v0.1. OpenJev-8B and OpenJev-1.7B are unchanged (v0.1).
• Data: alanhuangya/OpenJev-Data v2 adds:
  • parquet configs
  • new teacher and on-policy episodes
  • WebChain browser steps
  • wev-bench, a frozen browser lane that reports accuracy, calibration and latency together
  • redaction of personal data
• Inference: hybrid (Gated DeltaNet) backbones encode the page state once and reuse it for every question, so outputs are identical. Requests with oversized elements are shrunk instead of rejected.
• Training: resumable on any number of GPUs, and an on-policy (DAgger-style) collection pipeline.
• Package: pip install "wev-ai[serve,fast]" or pip install openjev-ai. Requires transformers 5. Pin a model version with wev.load(repo, revision=...).

Repository renamed from wev to OpenJev; old GitHub and Hugging Face links redirect.