Hermes-4.3-36B

Hermes-4.3-36B

Runs as a reasoning chat model that can think step by step or answer fast, and returns JSON that fits a schema you give it. Made for people who host their own language model.

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 is a reasoning chat model from NousResearch, built on the ByteDance Seed 36B base model. It can write out its thinking before it answers, or reply faster when you leave that off. It was also trained to return JSON that matches a schema you give it, and to fix broken JSON objects.

Who it suits: Developers and researchers who want an open reasoning model they can run on their own machine for math, code, logic and structured output.

Industry | Who uses it | How they use it | Example
Software | Developers | Run the model locally for code help and structured output | Ask for a JSON object that fits a schema
Education | Teachers and students | Work through math and logic problems step by step | Turn on reasoning mode for a maths question
Research | Machine learning researchers | Compare an open 36B model with other models | Run their own prompts and compare the answers

People who only want a quick chat answer can skip it and use a hosted chat instead.

Where to use it:

• Your own computer, after you download the files with Python on Windows, macOS or Linux
• Any program that loads models through the transformers library
• Nous Chat, the hosted chat the card points to

Use case | Who | What they get back
Math and logic problems | Students and researchers | An answer with the reasoning written out
Structured output | Developers | JSON that follows the schema you gave
Repairing JSON | Developers | A fixed version of a malformed object

How to set it up:

• Python with pip
• The transformers library, which the model is made for
• Disk space for about 72.3 GB of weights in BF16, or about 18.1 GB at 4-bit (an estimate)

1. Install the Hugging Face hub tool.
2. Download the model files to your computer.
3. Load them with the transformers library and send chats in the Llama-3 chat format.

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

How to use it: You send each chat with a system, user and assistant header, as the card shows below. To get reasoning mode, set the chat template flag thinking=True, or use the card's long system prompt that asks the model to think inside think tags.

<|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|>

Tool | What it does
thinking=True | Turns on reasoning mode in the chat template
think tags | Wrap the model's thoughts before its final answer

Example: You ask it to "Explain the photoelectric effect simply" in the chat format above. The model replies in the assistant turn with a short plain explanation. With reasoning mode on, it first writes its thoughts inside think tags, then the answer.

For newcomers and experts: If you are new, think of this as a chatbot brain you can download and run yourself. Start by downloading the files, then try one short question like the one above. The weights are large, so check your disk space first.

If you know the field, this is a 36.2B parameter model for transformers, based on Seed 36B. Post-training used about 5M samples and 60B tokens, mixing reasoning and non-reasoning data. The card says it is the first Hermes model trained in a decentralized way over the internet.

You send a chat in the Llama-3 format
Reasoning mode adds think tags when it is on
The model writes its thoughts, then the answer
You read the answer or parse its JSON

Good to know: The card's benchmark table lists these scores for the version trained over the internet.

Title: Hermes 4.3 36B scores from the card
OBQA: 96.6 points
MATH-500: 93.8 points
MMLU: 87.7 points
BBH: 86.4 points
DROP: 83.5 points
IFEval: 77.9 points
AIME 25: 69.3 points
SimpleQA: 6.0 points

• License: Apache 2.0, as the Hugging Face record lists it.
• Cost to run: no API key, account or paid service is named for a downloaded copy. You supply the hardware for the weights.
• Data it reaches: downloading the files reaches Hugging Face. The card does not describe the model reaching your files, the web or an account.
• Not for: the card names no banned uses. Its SimpleQA score is the lowest in its table, so check facts the model states.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 36.2B
• Weights: about 72.3 GB in BF16; about 18.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,559
• Likes: 321
• License: apache-2.0
• Languages: en
• Based on: ByteDance-Seed/Seed-OSS-36B-Base
• Quantized versions: 18
• Fine-tunes: 1
• Tags: transformers, safetensors, seed_oss, text-generation, instruct, finetune, reasoning, hybrid-mode, chatml, atropos, dataforge, roleplaying

Links

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

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