Hermes 4 405B is a large open-weight language model from Nous Research built on top of Meta's Llama-3.1-405B, positioned as a hybrid reasoning system that can interleave deliberate thinking with direct answers. Weights are publicly distributed in both BF16 and FP8 quantizations through the NousResearch Hugging Face organization, and a companion technical report (arXiv:2508.18255) documents the design. The post-training corpus was scaled substantially over Hermes the listed price, growing from roughly the listed price million samples and 1.2 billion tokens to about 5 million samples and 60 billion tokens blended across reasoning and non-reasoning data, with the aim of improving mathematics, coding, STEM reasoning, logic, creative writing, and adherence to requested output formats.
In practice the model is aimed at instruction following, conversational chat, tool and function calling, and reasoning-heavy workflows, and it supports JSON-mode structured output as well as streaming responses. The training emphasis on schema adherence and structured outputs makes it a good fit for agentic pipelines that need reliably parsed tool calls or JSON payloads, while the hybrid reasoning mode lets callers trade latency for deliberation on harder prompts. Hosted access is available through Nebius Token Factory using an OpenAI-compatible chat completions endpoint, with EU-routed alternatives also offered, making the model practical for teams that want frontier-class open weights without standing up their own inference infrastructure.
