OpenHermes 2.5

Teknium trained this 7 billion parameter chat model in late 2023 by fine-tuning Mistral-7B-v0.1 on roughly a million examples, most of them GPT-4 outputs plus filtered open datasets, all converted to ChatML with the Axolotl trainer. Compared with OpenHermes 2, the 2.5 run mixed in more code instructions, which raised HumanEval pass@1 from 43 percent to 50.7 percent and also lifted general scores such as TruthfulQA and AGIEval, at a small cost on BigBench. It honors system prompts across multi-turn chats and ships a chat template for the transformers apply_chat_template helper. The full weights take about 14.5 GB in safetensors, while community GGUF, GPTQ, AWQ, and EXL2 quantizations run on consumer GPUs or even CPUs through llama.cpp, LM Studio, or the openhermes tag in Ollama. The weights carry the Apache 2.0 license; teams with strict data provenance rules should weigh that the training set was largely produced by OpenAI models. The card has not changed since early 2024 and Nous Research has since shipped Hermes 3 and Hermes 4, so treat this as a small, well understood baseline for local chat or fine-tuning tests.