apus-ailab/APUS-Reflex-9B

Hugging Face
VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

APUS-Reflex-9B is a 9 billion parameter Qwen3.5-based decision model developed by APUS AI-LAB. It is specifically designed for browser action selection, workflow routing, and natural-language principle judgments, scoring dynamic candidates and returning their distribution. The model features a 32768 token context length and achieves 85.00% on the Frozen80 development panel, which includes tasks like Browser, HelpSteer3, BoolQ, MNLI, and attribute decisions.

Loading preview...

APUS Reflex 9B: A Specialized Decision Model

APUS-Reflex-9B, developed by APUS AI-LAB, is a 9 billion parameter model built upon the Qwen3.5 architecture. It is specifically engineered as a decision model, excelling in tasks such as browser action selection, complex workflow routing, and making natural-language principle judgments. This model is capable of scoring dynamic candidates provided with each request and outputting their probability distribution.

Key Capabilities

  • Decision-Making Focus: Optimized for selecting actions, routing workflows, and evaluating principles based on natural language.
  • Dynamic Candidate Scoring: Scores candidates supplied with each request and provides their distribution, useful for structured workflow outputs.
  • Qwen Language Representations: Leverages Qwen's robust language representations and vocabulary projection.
  • Performance: Achieves 68/80 (85.00%) on the Frozen80 development panel, which encompasses a range of tasks including Browser, HelpSteer3, BoolQ, MNLI, and attribute decisions.
  • Runtime Flexibility: Supports native runtime with effort=low (16 layers) for efficiency and effort=high (32 layers) for enhanced text generation.

Good For

  • Automated Browser Interactions: Ideal for applications requiring intelligent selection of browser actions.
  • Complex Workflow Automation: Suitable for routing and managing intricate workflows based on natural language instructions.
  • Principle-Based Judgments: Can be used for tasks that require evaluating and judging principles or attributes from text.
  • Structured Output Generation: Facilitates the assembly of decisions into structured workflow outputs.

Finetunes
4 models