Open-Jev-2B OpenVINO
Updated FP32 export using Optimum Intel's Qwen3.5 exporter, with merged Jev LoRA weights and the trained scalar decision head. This is a candidate scorer. Use the bundled runner for Jev requests; this is not a text-generation pipeline.
Validation status
All six saved PyTorch reference cases passed on CPU, including lengths 168 and 348 and unequal padded batches. Maximum absolute error: 0.00002575. Arc B580 GPU validation of this new export is pending. The previous custom export failed GPU parity at 168 tokens. This update is not yet a confirmed fix. See validation.json for versions, source revisions, and per-case results.
Run
hf download stijnvandoornum/Open-Jev-2B-OpenVINO --local-dir ./open-jev-ov
cd open-jev-ov
python -m pip install -r requirements.txt
python validate_export.py --device GPU
python run_model.py example_request.json --device GPU
python run_model.py --serve --model . --host 0.0.0.0 --port 8080 --device GPU
The model requires an OpenVINO-compatible Intel GPU driver for GPU execution. CPU inference is available with --device CPU. Compilation happens at load time. Tests cover numerical parity on six cases, not model quality or the full context limit. Original calibration and tokenizer are included. No quantization applied.
Sources: Open-Jev, Optimum Intel. See LICENSE, LICENSE-CODE, and UPSTREAM.md for attribution.