apus-ailab/APUS-Reflex-4B

Hugging Face
VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

APUS-Reflex-4B is a 4.5 billion parameter Qwen3.5-based decision model developed by APUS AI-LAB, gumpcheng, and zhangxu. It is specifically designed for browser action selection, workflow routing, and natural-language principle judgments, leveraging Qwen language representations. This model excels at scoring dynamic candidates and returning their distribution, making it suitable for applications requiring structured decision-making from natural language inputs.

Loading preview...

APUS Reflex 4B Overview

APUS-Reflex-4B is a 4.5 billion parameter decision model built upon the Qwen3.5 architecture, developed by APUS AI-LAB, gumpcheng, and zhangxu. It is specifically engineered for tasks involving browser action selection, complex workflow routing, and making natural-language principle judgments. The model reuses Qwen's robust language representations and vocabulary projection, allowing application code to assemble structured workflow outputs from its decisions.

Key Capabilities

  • Dynamic Candidate Scoring: The model can score dynamic candidates provided with each request and output their probability distribution, indicating relative preferences.
  • Structured Decision Making: It is designed to facilitate structured decision-making processes based on natural language inputs.
  • Flexible Runtime: The included native runtime supports both effort=low (16 layers) and effort=high (32 layers) modes, with the high-effort mode recommended for text generation tasks.
  • Performance: The merged model achieves 66/80 (82.50%) on the Frozen80 development panel, which includes tasks like Browser, HelpSteer3, BoolQ, and MNLI.

Good For

  • Automated Browser Interactions: Selecting appropriate actions within a browser environment based on contextual understanding.
  • Workflow Automation: Routing tasks and decisions within complex workflows using natural language instructions.
  • Principle-Based Judgments: Applying natural language principles to make judgments or classifications.
  • Applications Requiring Probabilistic Decision Outputs: When a distribution of preferences or likelihoods for various options is needed.

Finetunes
4 models