Nous Research drops Hermes 4 with hybrid reasoning and zero-fluff alignment

Nous Research just rolled out Hermes 4, a 70B open-weights model built on top of Llama-3.1-70B, along with a completely revamped Nous Chat web platform.

It is designed as a hybrid reasoning model. Instead of forcing you into a single rigid mode, Hermes 4 lets you toggle explicit deliberation using <think> tags. When activated, the model lays out its chain of thought before providing a clean final answer. If you just want quick answers, you can skip the internal reasoning and keep things fast.

What it does

The post-training data saw a massive overhaul. Nous Research expanded their dataset from 1 million samples and 1.2 billion tokens up to roughly 5 million samples and 60 billion tokens, heavily mixing verified reasoning chains with structured non-reasoning data. The result is noticeable gains across math, coding, STEM logic, and structured JSON generation, including the ability to repair broken schemas on the fly.

Tool calling is another standout feature. Hermes 4 supports structured function calling within a single assistant turn. The model can think through the prompt inside <think> tags, emit a clean <tool_call>, ingest the environment response, and continue deliberate reasoning before handing back the final response. Native tool parsers are already hooked up for inference engines like vLLM and SGLang.

Steerability is where the team drew a sharp contrast with closed-source giants. They introduced RefusalBench, a new benchmark evaluating how often models refuse helpful prompts across tricky scenarios. Hermes 4 scored at the top across both open and closed categories, ditching overzealous guardrails to faithfully follow whatever instructions and system personas you set.

How to run it

The full weights for Hermes 4 70B are up on Hugging Face using the standard Llama-3-Chat template. Running it in half precision requires enterprise hardware or multi-GPU setups, though quantized versions will fit more modest rigs. Default sampling parameters recommend a temperature of 0.6, top_p of 0.95, and top_k of 20.

If you do not want to manage local GPUs right away, the refreshed Nous Chat web app allows you to try the model directly in your browser. The technical report also notes formatting templates ready for Hermes 4 across both 70B and 405B scales.

Why it matters

Most frontier models are dialed in to be so safe they become stubborn, frequently refusing edge-case coding problems or complex creative prompts. Hermes 4 flips the priority back to developer control. It delivers high-tier chain-of-thought logic without the patronizing refusals. If you are building agentic workflows that rely on strict JSON and custom tool execution, this model offers one of the cleanest open foundations available right now.