Research

Building the future of edge AI with transparency and open research.

Maaza: Small Language Models for Edge Deployment

Maaza-MLM-135M-JSON-v1 and Maaza-SLM-360M-JSON-v1 are state-of-the-art small language models designed for structured JSON generation on edge devices.

Trained on the EdgeJSON benchmark, these models demonstrate that task-specific small models can outperform much larger general-purpose models on structured output generation.

Key Results

  • Maaza-MLM-135M: 55.1% JSONExact, 0.780 Field F1
  • Maaza-SLM-360M: 72.3% JSONExact, 0.878 Field F1
  • EdgeJSON Benchmark: 24 real-world schemas, 158 test examples
  • Deployment: CPU-only inference, <150MB disk footprint

Download Paper

Version 0.7 | 2025 | Metadata-scrubbed for privacy

Models & Benchmark

Citation

@article{maaza2025,
  title={Maaza: Small Language Models for Edge Deployment},
  author={CycleCore Technologies},
  year={2025},
  url={https://cyclecore.ai/research}
}