OpenHermes 2

A state-of-the-art 7B parameter model fine-tuned on 900,000 high-quality synthetic instructions to outperform base Mistral-7B.

OpenHermes 2 represents a significant leap in open-source LLM performance, built by teknium using a massive dataset of 900,000 entries. This dataset focuses on GPT-4 generated synthetic data, emphasizing code execution and complex reasoning tasks. By fine-tuning the Mistral-7B architecture, the model achieves high benchmarks across AGIEval and BigBench Hard, proving that curated data density can rival the capabilities of much larger proprietary systems. It remains a primary choice for developers seeking a fast, local model that handles nuanced instructions with precision.

https://huggingface.co/teknium/OpenHermes-2-Mistral-7B