Mesmer AI Detector 4B

Mesmer AI Detector 4B

Tells whether English text is human-written, AI-written or AI-edited and labels each sentence. For editors, teachers and researchers who want a signal, not proof, about how a text was made.

Made by mesmertech 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: Mesmer AI Detector 4B checks English text and says whether it looks human-written, AI-written or AI-edited, meaning human writing that an AI polished. It also labels each sentence. It is fine-tuned from Qwen3.5-4B-Base, and a bigger 9B sibling exists for higher accuracy.

Who it suits: It suits editors, teachers and researchers who want a signal about how a piece of English text was produced.

Industry | Who uses it | How they use it | Example
Publishing | Editors | Screen submissions for machine-written passages | Checking a freelance article sentence by sentence
Education | Teachers | Look at essays alongside other evidence | Reviewing a student essay for AI-edited passages
Research | Detection researchers | Compare detectors on public benchmarks | Re-running tests on Epoch-style imitation texts

You can skip it if your text is not in English, or if you need a verdict you can treat as proof, since it does flag some human writing.

Where to use it: It runs as a Python script or library on your own machine, with a GPU preferred. The card says CPU works but slowly, and it points to a hosted Space for trying it without a GPU.

• A command line, through the detector script
• Python code that imports the Detector class
• Windows, macOS or Linux with Python

Use case | Who | What they get back
Whole-text check | Editors | A human, AI-written or AI-edited verdict with probabilities
Sentence labelling | Teachers | Each sentence marked and shaded by the model
Benchmark testing | Researchers | Scores to compare against other detectors

How to set it up: You need Python and the libraries listed in the quickstart. An optional package adds faster GPU kernels, and without it the code falls back to plain PyTorch. Hugging Face lists about 8.4 GB in BF16.

• Python
• The libraries torch, numpy, tokenizers, safetensors, emoji and huggingface_hub
• Optional fla-core for faster GPU work
• A GPU is preferred

1. Install the required Python libraries.
2. Optionally install fla-core for faster kernels.
3. Download the detector script from the repository.
4. Run it with the model name and the text to check.

pip install torch numpy tokenizers safetensors emoji huggingface_hub
pip install fla-core   # optional: faster GPU kernels (otherwise pure PyTorch; CPU works too, but slowly)
hf download mesmertech/mesmer-ai-detector-4b mesmer_detector.py --local-dir .
python mesmer_detector.py --model mesmertech/mesmer-ai-detector-4b "Paste the text to check here."

How to use it: Run the script with your text, or load the Detector class in Python. Passing int8=True loads 8-bit weights that take about half the memory.

from mesmer_detector import Detector

det = Detector.from_pretrained("mesmertech/mesmer-ai-detector-4b")   # int8=True: 8-bit weights, about half the memory

Tool | What it does
mesmer_detector.py | Command line script that scores the text you pass in
Detector | Python class that loads the model for your own code

Example: You paste a paragraph into the command line script. It returns whether the text reads as human-written, AI-written or AI-edited, with a label for every sentence.

For newcomers and experts: If you are new, treat the result as a hint, not proof. Try it on a paragraph you wrote yourself and one from a chatbot to see how it labels each.

If you know the field, it is a 4.2B model with a three-way output. On the Open Pangram in-domain set it scores 0.918 ternary accuracy, close to the closed Pangram 3.2 at 0.920. The card says it flags more human writing than Pangram, for example 0.044 false positives on non-native English essays.

English text goes in
Model scores the whole text
Three-way label with probabilities
Each sentence labelled and shaded

Good to know: The card ranks it above every other open-weight detector it compared, and concedes a higher false-flag rate on human text.

• Cost to run: no API key, account or paid service is named, but a GPU helps a lot
• Data it reaches: the text you paste in, plus files downloaded from Hugging Face
• Not for: non-English text, commercial use, or treating a verdict as proof about a person

Title: In-domain ternary accuracy
Open Pangram: 0.895
Mesmer 4B: 0.918
Pangram 3.2 closed: 0.920
Mesmer 9B: 0.923

On Hugging Face, 13 November 2026

• Task: text classification
• Parameters: 4.2B
• Weights: about 8.4 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 68
• Likes: 3
• License: cc-by-nc-sa-4.0
• Languages: en
• Based on: Qwen/Qwen3.5-4B-Base
• Tags: safetensors, mesmer_ai_detector, ai-text-detection, ai-generated-text-detection, machine-generated-text-detection, ai-detector, text-classification

Links

• Model: https://huggingface.co/mesmertech/mesmer-ai-detector-4b
• Files: https://huggingface.co/mesmertech/mesmer-ai-detector-4b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mesmertech/mesmer-ai-detector-4b

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