# APUS-OpenJev-v1 · 4B

A decision model for browser agents and business workflows. This directory contains standalone BF16 weights and a runtime with selectable `effort="low"` and `effort="high"` compute budgets.

[Model family](../README.md) · [Architecture](../ARCHITECTURE.md) · [Runtime guide](RUNTIME.md)

## Quick start

Use a CUDA-capable PyTorch environment.

```bash
python -m pip install huggingface_hub
hf auth login
hf download apus-ailab/APUS-OpenJev-v1 \
  --include "4B-5949/*" --local-dir ./APUS-OpenJev-v1
cd ./APUS-OpenJev-v1/4B-5949
python -m pip install -r requirements.txt
python examples.py . --device cuda:0 --effort high
```

The included runtime provides compute-budget selection. Use `high` for text generation.

## Evaluation

With the full compute budget, this merged model scores **66/80 (82.50%)** on the [Frozen80 development panel](https://huggingface.co/datasets/apus-ailab/APUS-OpenJev-Eval-Frozen80): browser action selection, principle-based judgment, evidence-based questions, natural language inference, and attribute decisions.

This reused development panel is an engineering reference, not an independent benchmark. BF16 merging changes some candidate probabilities; decision thresholds require revalidation. See [evaluation results](merged-evaluation.json) and [runtime checks](evaluation/runtime-smoke.json) for details.

## Provenance

We thank the Qwen team for the [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) base model. Training details and source records are in [training.md](training.md); artifact hashes are in [release-manifest.json](release-manifest.json). Consult [LICENSE](LICENSE) and the [base-model](provenance/BASE-LICENSE.txt) and [source-project](provenance/SOURCE-PROJECT-LICENSE.txt) notices.

**Authors:** gumpcheng ([https://huggingface.co/xDAN2099](https://huggingface.co/xDAN2099)), zhangxu, [APUS AI-LAB](https://github.com/APUS-AI-Lab)