roskosmos19/Rhea-4B-fast-0409-high
Rhea-4B-fast-0409-high is a 4 billion parameter Qwen3ForCausalLM-based model developed by roskosmos19, optimized for agentic and coding tasks. This iteration focuses on faster, cheaper inference by adopting a single-pass reasoning style and a 32,768 token context window. It features a clean tool-calling format and optional tags for chain-of-thought, making it highly efficient for real-world agentic workflows.
Loading preview...
Rhea-4B-fast-0409-high: Optimized for Agentic & Coding Tasks
This model, developed by roskosmos19, is an optimized successor in the Rhea-4B-Coding / Athenea lineage, built on the Qwen3ForCausalLM architecture with 4 billion parameters. It is specifically engineered for enhanced performance in agentic and coding applications, prioritizing speed and cost-efficiency over its predecessors.
Key Optimizations & Capabilities
- Single-Pass Reasoning: Moves away from forced multi-pass generation, utilizing a single-pass approach with optional
<think>...</think>tags for chain-of-thought, significantly reducing latency and inference costs. - Efficient Context Handling: Features a 32,768 token context window, designed to cover typical real-world agentic workloads effectively.
- Clean Tool-Calling: Provides a reliable and streamlined tool-calling format, improving agentic readiness.
- Optimized Generation Defaults: Tuned with
temperature=0.4andtop_p=0.9for a balance of quality and speed, removing forced long outputs. - Lower Inference Cost: Achieves significantly lower VRAM and time consumption compared to previous versions due to its architectural changes.
- Strong Coding & Reasoning: Retains a core focus on robust coding and reasoning capabilities.
Ideal Use Cases
- AI Agents: Excellent for developing AI agents that require efficient tool use and flexible reasoning without unnecessary overhead.
- Code Generation & Analysis: Suited for tasks involving secure and efficient Python function generation, code validation, and general programming assistance.
- Cost-Sensitive Applications: Recommended for scenarios where inference cost and speed are critical factors, especially in production environments.
For optimal deployment, Q4_K_M or AWQ quantization is suggested, and keeping context within 16k–24k tokens in production is advised for peak performance.