APUS Reflex 4B

Branding note: This model family is now APUS Reflex. Some weight filenames, script/module names, served-model aliases, and the separate evaluation dataset ID retain their historical identifiers for download and runtime compatibility; these do not indicate different model versions.

English | 中文 · Collection · Model family · Technical Report · Runtime

A Qwen3.5-based decision model for browser action selection, workflow routing, and natural-language principle judgments. This repository contains 4B checkpoint-5949 merged BF16 weights, ready to download independently without a separate LoRA adapter.

Highlights

• Score dynamic candidates supplied with each request and return their distribution.
• The included native runtime supports effort=low (16 layers) and effort=high (32 layers). Use high for text generation.
• Reuse Qwen language representations and vocabulary projection; application code can assemble decisions into structured workflow outputs.

Quick start

python -m pip install huggingface_hub
hf download apus-ailab/APUS-Reflex-4B --local-dir ./APUS-Reflex-4B
cd APUS-Reflex-4B
python -m pip install -r requirements.txt
python examples.py . --device cuda:0 --effort high

Evaluation and training

The merged model in this repository scores 66/80 (82.50%) at full depth on the Frozen80 development panel, covering Browser, HelpSteer3, BoolQ, MNLI, and attribute decisions. See merged-evaluation.json and training.md. This reused engineering panel is not an independent blind benchmark or an end-to-end browser success rate.

Candidate probabilities express relative preference; calibration is required before interpreting them as correctness probabilities. BF16 merging changes some probabilities; see runtime evidence and limitations. The series 9B release selects checkpoint-3000, corresponding to the 85% development-panel result. Checkpoint-5949 remains available in the original family repository.

Series and downloads

The 4B, 9B, and 35B-A3B models use independent repositories grouped in a Collection. Hugging Face displays downloads per model. The original family repository and legacy paths remain available. The 35B-A3B merged release is subject to its own validation and publication status.

GGUF and MLX versions

Quantized versions of this model (Frozen80 66/80) for Ollama / llama.cpp / LM Studio and Apple Silicon Macs: GGUF collection · MLX collection.

ollama run hf.co/apus-ailab/APUS-Reflex-4B-GGUF:Q8_0 --think=false

Ollama needs thinking disabled (--think=false, or "think": false in the API). Per-question results and usage are in each repository.

License and acknowledgments

We thank the Qwen/Qwen3.5-4B team. See LICENSE, provenance, and the pinned source and file identities in release-manifest.json.

Authors: gumpcheng (xDAN2099), zhangxu, APUS AI-LAB.