Hermes 4 405B

Hermes 4 405B is Nous Research's flagship 405 billion parameter model, introducing hybrid-mode reasoning with explicit <think>…</think> segments that the model uses when it decides to deliberate, letting users trade speed for depth. It targets stronger performance across math, code, STEM, logic, and creative tasks, trained on roughly 5 million post-training samples spanning about 60 billion tokens with an emphasis on verified reasoning traces. Key differentiators include state-of-the-art RefusalBench results for being helpful while conforming to the user's values without censorship, schema-faithful JSON generation with self-repair of malformed objects, and broadly neutral alignment that can be steered toward user preferences. Hermes 4 expanded its training corpus roughly five-fold compared to Hermes 3, which Nous Research links to improved steerability and substantially lower refusal rates, while retaining function calling, structured outputs, and multi-turn dialogue. With a 131,072 token context window and OpenAI-compatible APIs, Hermes 4 405B targets teams building systems that need both reasoning depth and values alignment, from complex problem-solving to creative generation where user personalization matters.