Hermes 4.3 36B is a 36-billion-parameter model from Nous Research, built for open conversation and role-play. According to its developers, it rarely refuses. The template runs it as a GGUF file on a 32 GB card such as an RTX 5090, with a chat in the browser and an API behind a protected link.
In short
Hermes 4.3 36B runs on cards with 32 GB or more. The most affordable suitable card right now is the Tesla V100 32 GB from 23 ₽ per hour, billed by the minute.
Updated 13 October 2026
The model needs a 32 GB card: the Q4_K_M GGUF file is about 22 GB, and the context and buffers take the rest. An RTX 5090 with 32 GB fits. For long sessions with a large context, an RTX A6000 with 48 GB leaves more room.
Prices and availability come from the live catalog and update automatically.
Hermes keeps a conversation going well and holds a role when the character is described up front: name, temperament, way of speaking, facts about the world. It talks about open topics easily and discusses books and ideas. For code and analysis, Qwen 3.8 is usually the more convenient choice; for role-play scenes and open chat, pick Hermes.
GGUF is a format for compressed models. The Q4_K_M quantization stores the weights at roughly 4 to 5 bits per parameter instead of 16, so 36 billion parameters fit in a file of about 22 GB. Answers lose a little quality compared with the full model, but the memory needs drop a lot. For how to size the memory, see the VRAM guide for LLMs.
The download from Hugging Face takes a few minutes on first start. After that the model stays in memory while the server runs.
Set the character with a system message: in the API it is the first message, with the role system; in the chat it goes into the system prompt field. An example request with a character and a user line:
curl "https://<id>-11434.run.gpu.nz/v1/chat/completions" \
-H "Authorization: Bearer <token>" \
-H "Content-Type: application/json" \
-d '{"model": "hf.co/NousResearch/Hermes-4.3-36B-GGUF:Q4_K_M", "messages": [{"role": "system", "content": "You are Lea, captain of a cargo ship. You speak briefly and with irony."}, {"role": "user", "content": "What do we do about the pirates at the asteroid belt?"}]}'The context is 16,384 tokens. Over a long session, older lines stop counting, so repeat the key details of the character in the system message. If several people call the API at once, look at the vLLM template.