O
Business AI
OpenJev
Run Jev-style typed decisions locally: in-browser or on a 3090
4,881 stars
MIT
last push 2026-09-23
OpenJev reproduces the Jev interface pattern with open models: read typed option probabilities directly from a model's logits - no answer sentence, JSON repair, or decoding loop. On a single RTX 3090 with frozen Qwen3.5-4B, 21 decisions complete in 1.0 seconds with zero output tokens (5.2x faster than autoregressive JSON, identical results), and shared-state reuse pushes throughput to 20 decisions/second on 777-decision workloads. On the 102-row TypeSafe subset, the open 4B model reaches 0.845 balanced accuracy against Jev's published 0.883. Ships as a browser-only WebGPU demo (no waitlist, no backend) and a Python scorer with pinned model revisions, prompt hashes, and raw results with checksums. The repo states plainly it does not reproduce Jev's undisclosed model or training.
daily snapshots
2026-09-22 to 2026-10-24
At a glance
Primary use case
Same judgments as Jev, on your own hardware: batch classification of records, content moderation screens, agent step decisions (retry / route / pause), form routing, dataset tagging. The browser demo needs no install and no API key - 21 questions in one second on a 3090. Best fit when you want typed decisions without sending code or data to a third party.
Who's behind it
TheoLeeCJ (independent)
- License
- MIT
- Platforms
- browser, cli
- Install
pip install -e '.[test]'
- Model support
- Open models: Qwen3 0.6B, MiniCPM5 2B, Qwen3.5 4B (GGUF browser / BF16 CUDA)