Hermes 3: An Uncensored 70B Model That Runs on Two GPUs

Running a 70B model at home usually means hardware you do not own. Hermes 3 is worth a look anyway, because the hardware it asks for is smaller than you would expect.

Two 24GB cards at Q4 or FP4 will do it, and the context window is a native 128k tokens. That is a serious open-weight model on a two-GPU desk.

What Happened

Hermes 3 (Llama 3.1 70B) comes from Nous Research, and the team built it around a philosophy they call neutral alignment. The goal is a model that follows the prompt you give it rather than refusing it.

Nous designed it for three things: very close compliance with what the user asks, autonomous agentic reasoning, and complex tool-calling. It ships with no built-in censorship bias.

The specs the source lists:

• VRAM: 2x 24GB for Q4 or FP4, or 48GB for FP8
• System RAM: 64GB
• Storage: about 45GB
• Context window: 128k tokens, native

Why It Matters

Those VRAM numbers are the interesting part. Two 24GB cards is a workstation, not a datacenter rack, so this is a model you can realistically run yourself instead of renting it by the hour. The 128k native context means long documents and long agent sessions fit without chopping them into pieces first.

Tool-calling is what makes a model useful inside an agent loop, and it is one of the three things Nous aimed at here. If you are wiring a model into something that has to call functions and keep going on its own, that focus is the reason to care. A 70B open-weight model with 128k context and no refusal layer is a specific thing, and this one is built for exactly that.