APUS-OpenJev-v1-4B-GGUF

APUS-OpenJev-v1-4B-GGUF

APUS-OpenJev-v1-4B is a Qwen3.5-4B-based decision model from APUS AI-LAB, designed for browser action selection, workflow routing, and natural-language principle judgments rather than open-ended text generation — this repository provides the 4B checkpoint-5949 merged BF16 weights, downloadable independently without a separate LoRA adapter. It scores dynamic candidates supplied per request and returns their preference distribution, reusing Qwen's language representations and vocabulary projection so application code can assemble decisions into structured workflow outputs; its native runtime supports two effort levels — low (16 layers) and high (32 layers, recommended for text generation). On the internal Frozen80 development panel (covering Browser, HelpSteer3, BoolQ, MNLI, and attribute-decision tasks), the merged model scores 66/80 (82.50%), though the authors note this is a reused engineering panel rather than an independent blind benchmark or end-to-end browser success rate, and that candidate probabilities express relative preference rather than calibrated correctness (with BF16 merging further shifting some probabilities). It's part of a three-size family (4B, 9B, 35B-A3B) hosted in separate repositories under a shared collection, with the 9B release instead selecting checkpoint-3000 (85% on the same panel), and is released under Apache 2.0 as a finetune of the Qwen3.5-4B base.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp