Hermes 4 70B

Hermes 4 70B is a hybrid-mode reasoning model based on Llama 3.1 70B, introducing optional <think>…</think> deliberation segments that can be activated through the system prompt. This lets developers trade inference speed for reasoning depth on demand, while keeping function calling and structured output capabilities within assistant turns. Training used roughly 5 million post-training samples across about 60 billion tokens blended between reasoning and non-reasoning data, emphasizing verified reasoning traces and gains in math, code, STEM, logic, and creativity. Compared with Hermes 3 70B, Hermes 4 improves schema adherence for JSON generation, reduces refusal rates, and adopts broadly neutral alignment that can be steered without censorship. With a 131,072 token context window and OpenAI-compatible APIs, Hermes 4 70B balances capability and compute for production deployments that need both reasoning and tool use. It suits teams building agentic systems where controlled reasoning improves quality, code-generation pipelines needing multi-step logic, and applications where developer-preferred alignment matters.