Nanthasit/sakthai-context-1.5b-merged

SakThai Context 1.5B Merged: Tool-Calling Agent Model

Nanthasit/sakthai-context-1.5b-merged is a 1.54 billion parameter model, fine-tuned from Qwen2.5-1.5B-Instruct, specifically designed for tool-calling and agentic tasks. It is the most downloaded model in the SakThai family, optimized for scenarios where an LLM needs to interact with external tools.

Key Capabilities & Features

• Tool-Calling Optimization: Fine-tuned to understand when to call tools versus answering directly, using a structured <tool_call> XML output format.
• Multi-turn Conversations: Maintains context across multiple turns, enabling complex agentic workflows.
• Full-Weight Checkpoint: Merged from QLoRA training, eliminating the need for PEFT during deployment.
• Structured Output: Emits tool calls in a specific XML format, requiring a <tools> block in the system prompt for reliable function calling.
• Efficient Inference: Achieves approximately 5.23 tokens/sec on CPU with q4_k_m GGUF quantization, making it suitable for edge devices and local-first applications.
• Extensive Context Window: Supports a 32,768 token context, though performance may degrade at the upper limits.

Ideal Use Cases

• AI Agents: Building autonomous agents that require tool interaction for task completion.
• Tool-Integrated Applications: Developing applications where the model needs to trigger specific functions or APIs.
• Resource-Constrained Environments: Its small size and CPU-friendly inference make it suitable for deployment on devices with limited computational resources.
• Structured Decision Making: Excels in tasks requiring structured outputs and clear decision-making for tool utilization.