APUS-OpenJev-v1-4B
APUS AI Lab's decision model for browser action selection, workflow routing and principle judgments. A Qwen3.5-4B fine-tune that scores 2 to 16 candidates supplied with each request, or a Yes/No proposition, from label-token logits, at a selectable depth of 16 or 32 layers.
Candidates get single-token labels A to P and the answer is read at the answer boundary, with nothing sampled; noul uses fixed Yes/No candidates. APUS's runtime (pinned to transformers 5.16.1, CUDA) runs 16 layers (low effort) or all 32 (high); inputs over 8,192 tokens are rejected, not truncated. The family repo has a vLLM launcher and a gateway compatible with the TypeSafe SDK, including an ordinal Score that APUS says is not validated. Probabilities are not calibrated, and APUS says the BF16 merge shifted some of them. On APUS's own 80-question development panel (public, but used for model selection, so not blind) the 4B answers 82.5% at full depth; the 9B 85.0% and the 35B-A3B 88.75%. Siblings APUS-OpenJev-v1-9B and 35B-A3B, plus APUS's own GGUF and MLX builds. Training-data licences are mixed and the data is not released.
What it decides
- choice — picks one option from a set
- noul — answers a yes/no question with one probability
- route — sends the input to one of several destinations
At a glance
| Parameters | 4.5B |
| Base model | qwen/qwen3.5-4b |
| Maker | APUS AI Lab |
| Released | 2026-09-22 |
| License | apache-2.0 |
| Reported accuracy | 82.5% |
| Reported latency | 25.58 ms p50 / 222.36 ms p95 full HTTP response for the 9B on one RTX PRO 6000 Blackwell (vLLM) |
Get the weights
pip install systemonemodels
systemone pull apus-ai-lab/apus-openjev
The files are served from the maker's Hugging Face repository, apus-ailab/APUS-OpenJev-v1-4B, and verified against the checksums recorded here.