roskosmos19/Orca-2-4B-hight

Orca-2-4B-hight: Optimized for Agentic & Coding Tasks

This model, based on the Qwen3ForCausalLM architecture with 4 billion parameters, is an optimized successor in the Athenea/Rhea coding lineage. It is specifically engineered for agentic and coding tasks, prioritizing faster and cheaper inference while maintaining strong performance.

Key Differentiators & Capabilities

• Efficient Reasoning: Employs a single-pass reasoning style with optional <think>...</think> tokens, allowing the model to decide when chain-of-thought is beneficial, unlike previous multi-pass approaches.
• Optimized for Agents: Features an improved tool-calling template and a clean, lean set of special tokens, making it highly suitable for agentic workflows.
• Cost-Effective: Significantly reduces inference cost and latency by removing forced long outputs and artificial multi-pass overhead.
• Strong Coding Focus: Retains a strong emphasis on coding and reasoning, encouraging precise, secure, and efficient solutions through its system prompt.
• Generous Context Window: Offers a 32,768 token context length, ample for real-world agentic workloads.

When to Use This Model

This model is ideal for developers building applications that require:

• Agentic AI systems needing reliable tool-calling and efficient reasoning.
• Code generation and problem-solving where speed and cost are critical.
• Applications benefiting from a large context window without the overhead of forced long generations.

It is recommended to use settings like temperature: 0.4 and top_p: 0.9 for optimal quality and speed, and quantizations like Q4_K_M or AWQ for best performance.