Osmosis-Structure-0

Osmosis-Structure-0 is a 600-million parameter small language model optimized for high-precision JSON extraction and structured data generation.

Osmosis-Structure-0 (specifically the 0.6B variant) is a compact, high-efficiency SLM built on the Qwen3 architecture and refined through reinforcement learning on 500,000 JSON-to-natural language pairs. It excels at post-processing unstructured reasoning traces into clean, schema-compliant formats, making it a powerful tool for agentic workflows and local deployments. By focusing on key-value accuracy rather than generalist knowledge, this model delivers reliable structured outputs on consumer-grade hardware (requiring only 1.2 GB of storage) while outperforming much larger models on mathematical and logical extraction benchmarks.

https://huggingface.co/osmosis-ai/Osmosis-Structure-0.6B