Nanthasit/sakthai-context-0.5b-merged

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Nanthasit/sakthai-context-0.5b-merged is a 494 million parameter Qwen2.5-0.5B-Instruct based model, fine-tuned for tool-calling with structured XML output. Designed for extreme edge deployment, it runs in approximately 1 GB RAM on devices like Raspberry Pi, old laptops, or phones. This model excels at reliable tool selection, making it ideal for resource-constrained applications requiring function calling capabilities.

Loading preview...

Overview

Nanthasit/sakthai-context-0.5b-merged is a highly accessible 494 million parameter model, built upon the Qwen2.5-0.5B-Instruct architecture. It is specifically fine-tuned for tool-calling, generating structured XML output for function selection. A key differentiator is its minimal resource footprint, requiring only ~1 GB RAM, making it suitable for deployment on edge devices such as Raspberry Pi, older laptops, or mobile phones.

Key Capabilities

  • Edge Deployment Ready: Optimized to run efficiently on low-resource hardware.
  • Tool-Calling: Fine-tuned to accurately select tools, verified with 5/5 accuracy in llama.cpp evaluations.
  • Structured Output: Emits tool calls in a structured <tool> XML format, with arguments in JSON.
  • Quantization: Available in GGUF Q4_K_M quantization (379 MB) for further memory efficiency.
  • Context Window: Features a substantial 32,768 token context window, inherited from its base model.

Limitations and Considerations

While highly efficient for tool selection, this 0.5B model has limitations in deep reasoning or complex multi-step planning. Tool argument construction has not been systematically benchmarked and may be less reliable than tool selection. It requires the <tools> XML block in the system prompt for tool-calling behavior and is primarily English-only. Benchmarks are internal and not independently verified. For higher reasoning or argument-filling reliability, the 1.5B Merged variant is recommended.

Good For

  • Applications requiring reliable tool selection on resource-constrained devices.
  • Edge computing scenarios where minimal RAM usage is critical.
  • Developers building lightweight assistants or agents for embedded systems.

Finetunes
1 models