Hermes-4.3-36B-GGUF

Hermes-4.3-36B-GGUF

A 36B hybrid reasoning model in GGUF form that can think before it answers, write valid JSON and follow your system prompt. For people who run a large model on their own machine.

Made by NousResearch on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Hermes-4.3-36B-GGUF holds GGUF files of Hermes 4.3 36B, a hybrid reasoning model from Nous Research built on the ByteDance Seed 36B base. It can think in a visible block before answering, or answer straight away. It was trained to produce valid JSON for a given schema, and the authors call it their first Hermes model trained in a decentralized way over the internet.

Who it suits: It suits people who want a large open reasoning model on their own hardware, for math, code, structured output and general chat.

Industry | Who uses it | How they use it | Example
Software | Developers | Get valid JSON that fits a schema | Ask for output that matches a given structure
Research | Analysts | Work through math, logic and science questions | Switch on reasoning for a hard problem
Education | Students | Ask for plain explanations | Ask for a simple account of a physics idea
Writing | Writers | Draft text in a tone you set | Describe the style in the system prompt

If your machine cannot hold a model of this size, you can skip it, since even the 4-bit weights are about 18.1 GB.

Where to use it: You run the GGUF files in a local app or library that reads that format, on Windows, macOS or Linux. Python with the huggingface_hub package handles the download.

• An app or library that reads GGUF files
• Windows, macOS or Linux, with memory for about 18.1 GB of weights
• Python with the huggingface_hub package, for the download

Use case | Who | What they get back
Reasoning questions | Students and analysts | Answers preceded by a visible think segment when reasoning is on
Structured output | Developers | JSON that fits a given schema, plus repair of broken JSON objects
Everyday chat | General users | Quicker direct answers when you leave reasoning off

How to set it up: You need Python, the Hugging Face hub package and an app that reads GGUF files. Hugging Face estimates about 18.1 GB for the weights at 4-bit. The part of the card that explains the files is not shown here.

• Python on Windows, macOS or Linux
• The Hugging Face hub package
• Enough memory for the weights

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Open a GGUF file in your app and pick the Llama-3-Chat prompt format.

pip install -U huggingface_hub
hf download NousResearch/Hermes-4.3-36B-GGUF

How to use it: The model uses the Llama-3-Chat format, with role headers around the system, user and assistant turns. This is the card's basic chat layout.

<|start_header_id|>system<|end_header_id|>

You are Hermes 4. Be concise and helpful.<|eot_id|>
<|start_header_id|>user<|end_header_id|>

Explain the photoelectric effect simply.<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>

To turn on reasoning mode, set the thinking flag in the chat template, or use the reasoning system prompt that the card gives.

Tool | What it does
thinking=True | Turns on reasoning mode in the chat template
think tags | Wrap the model's reasoning before its answer
System prompt | Sets the model's role, as in the card's chat layout

Example: You write a system prompt asking for short, helpful answers, then ask for a plain explanation of the photoelectric effect. The model replies in the assistant turn, with a think segment first if reasoning is on.

For newcomers and experts: If you are new to this, GGUF is a file format for running a model on your own machine, and hybrid reasoning means it can think first or answer straight away. A first small step is to download the files and ask one short question with reasoning off.

If you know the field, the post-training set grew from about 1M samples and 1.2B tokens in Hermes 3 to about 5M samples and 60B tokens. Training stressed verified reasoning traces and answers that keep to the requested format.

A Llama-3-Chat prompt goes in
The model decides whether to think first
It writes a think segment, then the answer
The reply comes back

Good to know: These are the Psyche-trained version's scores from the card's table, picked from twelve benchmarks.

Title: Benchmark scores for Hermes 4.3 36B Psyche
OBQA: 96.6
MATH-500: 93.8
MMLU: 87.7
BBH: 86.4
DROP: 83.5
MMLU-Pro: 80.7
IFEval: 77.9
AIME 24: 71.9

• License: Apache-2.0
• Limits: the card lists English, and its SimpleQA score is only 6.0, so check any facts it gives
• Refusals: the authors tuned it to refuse less often and call its alignment broadly neutral, so add your own safeguards
• Cost to run: no API key or paid service is named, and the 4-bit weights are about 18.1 GB
• Data it reaches: the download comes from Hugging Face, and the model works on the prompts you give it
• Not for: small machines, or facts you cannot check

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 36.2B
• Weights: about 18.1 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 28,565
• Likes: 99
• License: apache-2.0
• Languages: en
• Based on: ByteDance-Seed/Seed-OSS-36B-Base
• Tags: transformers, gguf, instruct, finetune, reasoning, hybrid-mode, chatml, atropos, dataforge, roleplaying, chat, text-generation

Links

• Model: https://huggingface.co/NousResearch/Hermes-4.3-36B-GGUF
• Files: https://huggingface.co/NousResearch/Hermes-4.3-36B-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download NousResearch/Hermes-4.3-36B-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.