Hermes 2 Pro Llama 3 8B

Built on the Llama 3 8B foundation as an upgraded version of Nous Hermes 2, this model introduces specialized tokens for function calling and JSON structured output that stream as single tokens, enabling cleaner parsing during generation. The design uses a custom system prompt paired with a new chatml role structure to support multi-turn function calling interactions reliably. The architecture targets agentic applications where models need to invoke external tools, return well-formed data, and maintain coherent conversation across tool-use cycles.

Training combines an updated and cleaned OpenHermes 2.5 Dataset with a proprietary in-house Function Calling and JSON Mode dataset developed by Nous Research in collaboration with interstellarninja and Fireworks.AI. The model achieves strong evaluation scores on structured tasks, reaching 90% on function calling benchmarks and 84% on structured JSON output evaluations. These capabilities make it well-suited for coding assistance, automated data extraction, and multi-step tool orchestration in agentic pipelines.