Openhermes2.5 Mistral 7B

OpenHermes 2.5 Mistral 7B is a 7 billion parameter fine-tune of Mistral-7B-v0.1 trained on around 1 million primarily GPT-4 generated entries plus other high-quality open datasets, with the notable addition of 7 to 14 percent code instruction data. This code training also lifted non-code benchmarks, while improving HumanEval from 43% to 50.7% pass@1. The model uses the ChatML prompt format, enabling structured multi-turn dialogue with system prompts. Reported scores include a GPT4All average of 73.12, AGIEval average of 43.07%, BigBench average of 40.96%, and TruthfulQA of 53.04. Quantized variants in GGUF, GPTQ, AWQ, and EXL2 formats are available for efficient deployment. OpenHermes 2.5 suits developers building applications where 7 billion parameters are enough, particularly code-heavy workloads that benefit from its blended training, or scenarios where a small, fast model keeps compute cost down. It represents an earlier point in the Hermes lineage, before Hermes 2 Pro's agentic enhancements and the Hermes 3 and 4 reasoning additions.