roskosmos19/Rhea-4B-Coding

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Rhea-4B-Coding by Roskosmos19 is a 4 billion parameter model based on the Qwen3 architecture, optimized for code reasoning, debugging, and multi-pass problem solving. It specializes in detailed 3-pass reasoning for software development, algorithm design, and code comprehension tasks, utilizing explicit reasoning traces. Fine-tuned on high-quality programming data, this model excels at iterative code refinement and agentic tool integration.

Loading preview...

Model Overview

Rhea-4B-Coding, developed by Roskosmos19, is a 4 billion parameter model built on the Qwen3 architecture. It is an optimized version of Aquiles-ai/Athenea-4B-Coding, specifically designed for advanced code reasoning, debugging, agentic tools, and multi-pass problem solving. The model is trained on high-quality programming data, including explicit reasoning traces, to facilitate a detailed 3-pass reasoning process for software development and algorithm design.

Key Capabilities

  • Multi-Pass Processing: Employs a unique 3-step reasoning process using <think>, <review>, and <final> tags for initial implementation, self-review, and final optimization.
  • Agentic Tools: Designed to integrate with AI agents for enhanced problem-solving.
  • Step-by-step Code Reasoning: Generates detailed reasoning within thinking blocks.
  • Self-review: Capable of detecting bugs and optimizing code iteratively.
  • Specialization: Excels in algorithmic and debugging tasks, improving logical consistency.
  • Uncensored Output: Provides full expressive freedom for research and experimentation.

Training and Usage

The model was fine-tuned using the Aquiles-ai/Athenea-Coding-100k dataset, which contains diverse programming challenges and structured reasoning chains across multiple languages. It is compatible with open inference frameworks like Transformers and vLLM, with recommended multi-pass inference for optimal results. The model configuration emphasizes full multi-pass output capacity and faster inference for long outputs.