Hermes 4 14B

Hermes 4 14B

A 14B hybrid reasoning model from Nous Research, built on Qwen3-14B. It can think step by step, call tools, and return JSON that follows a schema. It uses the ChatML prompt format.

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 14B is a hybrid reasoning model from Nous Research, built on Qwen3-14B. It can write out its thinking between think tags when it decides to deliberate, or answer directly when you want speed. It was trained to produce valid JSON for a schema you give, to repair malformed objects, and to call tools inside a single reply. It uses the ChatML prompt format.

Who it suits: It suits developers who want an open 14B model with optional reasoning, tool calls and structured output, and researchers who study how system prompts steer a model.

Industry | Who uses it | How they use it | Example
Software development | App developers | Build assistants that call tools and return JSON | Give it a weather tool and ask about a city
Data work | Teams that extract structured data | Get output that matches a schema | Ask for JSON in your schema and have it repair a broken object
Research | Researchers of steerability | Test how system prompts change style and policy | Add system instructions and compare the responses

If you want a model with strict built-in refusals, you may prefer another one. The card describes lower refusal rates and a broadly neutral alignment.

Where to use it:

• Python projects with the transformers library, as the Hugging Face page lists
• Local runners and servers that accept the ChatML format
• Agent and tool-calling setups that parse tool_call tags
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Reasoning | Developers | Thinking between think tags, then a final answer
Tool calling | Agent builders | Tool calls inside one reply, then an answer based on the tool response
Structured output | Data teams | JSON that follows the schema you supply

How to set it up: What you need first:

• Python and the Hugging Face command line
• A computer with a GPU and enough memory for a 14B model, or one of the quantized copies other people published
• A runner that applies the model's chat template

1. Install the Hugging Face tools.
2. Download the model files.
3. Load the model with transformers and send prompts in the ChatML format.

pip install -U huggingface_hub
hf download NousResearch/Hermes-4-14B

How to use it: The model reads ChatML, with role headers and special tags. Reasoning mode turns on through the chat template flag thinking=True, or through the card's system prompt. You can add your own system instructions before or after it, and they change the policies, style, effort of thinking and format of the answer.

<|im_start|>system

You are Hermes 4. Be concise and helpful.<|im_end|>
<|im_start|>user

Explain the photoelectric effect simply.<|im_end|>
<|im_start|>assistant

Part | What it does
thinking flag | Turns on reasoning mode through the chat template
keep_cots flag | Keeps the content between the think tags
System prompt | Adjusts policies, style, effort and format
Tool tags | The tools, tool_call and tool_response tags mark function calls in one turn

Example: You send the card's basic prompt and ask for a simple explanation of the photoelectric effect. With reasoning on, the model may write its thoughts between think tags first, then give the final answer.

For newcomers and experts: A good first try is the basic chat prompt above. Then switch on reasoning mode with thinking=True and compare the two answers.

For experts, the training used a newly synthesized post-training corpus that stresses verified reasoning traces. The model deliberates only when it chooses to, and it can call a function after its reasoning, within the same assistant turn. After a tool responds, it continues reasoning inside the think block and writes the final answer. You can combine the tool definition message with the reasoning system prompt in one setup.

You send a ChatML prompt with a system message
Reasoning mode may add a think segment
The model can call a tool inside the same turn
It reads the tool response and writes the answer

Good to know:

Title: Post-training samples in millions
Hermes 3: 1 million
Hermes 4: 5 million

• Training data: the post-training set grew from about 1 million samples and 1.2 billion tokens to about 5 million samples and 60 billion tokens, mixing reasoning and non-reasoning data.
• Refusals: the card reports lower refusal rates, describes a benchmark its authors made called RefusalBench, and calls the alignment broadly neutral. Check this against your own policies and legal duties.
• Benchmarks: the card shows benchmark images and points to a technical report, so check them for numbers.
• Variants: the page lists many quantized and fine-tuned versions made by other people.
• License: apache-2.0
• Cost to run: no API key or paid service named. The card also mentions a hosted chat from Nous Research, which is separate.
• Data it reaches: it downloads files from Hugging Face, and a local run stays on your machine.
• Not for: places where you need strict refusals built into the model.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 425K
• Weights: about 1 MB in BF16; about 1 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 6,062
• Likes: 205
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3-14B
• Quantized versions: 31
• Fine-tunes: 9
• Tags: transformers, safetensors, qwen3, text-generation, qwen-3-14b, instruct, finetune, reasoning, hybrid-mode, chatml, atropos, dataforge

Links

• Model: https://huggingface.co/NousResearch/Hermes-4-14B
• Files: https://huggingface.co/NousResearch/Hermes-4-14B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download NousResearch/Hermes-4-14B

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