MANTA-27B

MANTA-27B

Scores the answer options of medical questions from text or images in one pass, without writing text. A research model from Stanford's Gevaert Lab, not a medical device.

Made by gevaertlab 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: MANTA-27B is a medical decision model from the Gevaert Lab in Stanford University's Department of Computational Medicine. You give it a medical image or text, plus questions with answer options. Each option of each question gets its own probability from a single run through the network, and no text is written.

Who it suits: It suits researchers who study medical question answering and want calibrated option probabilities. The card calls it a research model, not a medical device.

Industry | Who uses it | How they use it | Example
Medical AI research | Computational medicine groups | Score answer options for medical questions in a single pass | Run exam-style questions with text or images and read the probabilities
Health technology | Teams prototyping decision support | Test it on their own data, with clinicians reviewing the results | Compare its probabilities with expert answers on a private set
Medical education research | Teams studying exam questions | Benchmark against question sets such as MedQA and MedMCQA | Measure accuracy across a set of multiple-choice questions

You can skip it if you need a model that writes text, makes patient decisions, or runs on a machine without a large GPU.

Where to use it:

• A machine with one GPU that has about 60 GB of memory
• Python with the transformers library
• Windows, macOS and Linux with Python, to download the files
• Research evaluation scripts that feed it questions and options

Use case | Who | What they get back
Medical question scoring | Researchers | A probability for each answer option
Benchmarking | Evaluation teams | Accuracy numbers to compare with other models
Image and text questions | Medical AI groups | Option probabilities from either an image or text input

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• A GPU with about 60 GB of memory, since the BF16 weights are 55.0 GB
• Disk space for the weights, which come as 12 safetensors shards

1. Install or update the Hugging Face download client.
2. Download this repository's files to your computer.
3. Open the model card for the loading and running steps, which are not in the text available here.

pip install -U huggingface_hub
hf download gevaertlab/MANTA-27B

How to use it: You pass in a medical image or text, a set of questions and the answer options for each. Every option comes back with its own probability, from one run through the model.

Tool | What it does
pip install -U huggingface_hub | Installs or updates Hugging Face's download client
hf download | Fetches a model repository's files to your computer

Example: You pass in a multiple-choice medical exam question as text, with four answer options. The model returns four probabilities, one per option, and no written explanation. A clinician should still check any result.

For newcomers and experts: MANTA is not a chatbot. It picks among options you supply and says how likely each one is. A good first step is to read the model page, then fetch the files with the two commands above.

For readers who know the field, it takes Qwen3.8-27B, a vision-language model, as its base and attaches a freshly trained decision head that scores each option directly. The language and vision towers are both tuned end to end so the probabilities stay calibrated. Training used about 8.5 million decision requests, 68 percent of them medical, and the card reports an ECE-15 of 0.021 in-domain.

Provide a medical image or text
Add questions and answer options
The backbone encodes the input
The decision head scores each option
Probabilities come back per option

Good to know:

• Research use: the card says do not use outputs for diagnosis, treatment or patient decisions without independent validation and review by clinicians.
• Sibling model: MANTA-4B needs about 16 GB of GPU memory, has a 68 ms median latency and scores 85.3 percent on the card's medical benchmark.
• This release: about 60 GB of GPU memory, a 148 ms median latency and 86.0 percent on the same benchmark.
• Results: the figures come from the authors' own evaluation, so check them on your own data.

Title: Accuracy on the card's 6,700 medical decision questions
MANTA-27B: 86.1%
Gemini 3.8 Flash: 76.0%
Claude Sonnet 5.5: 74.6%
GPT-6 Luna (chat): 62.4%
MedGemma 27B (local): 52.3%

• License: Apache 2.0.
• Cost to run: none named. You need a GPU with about 60 GB of memory.
• Data it reaches: the medical text or images you pass in. The text available here names no network use.
• Not for: diagnosis, treatment or any patient decision without independent validation, and not for writing text.

On Hugging Face, 13 November 2026

• Task: text classification
• Library: transformers
• Parameters: 27.5B
• Weights: about 55 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 46
• Likes: 4
• License: apache-2.0
• Languages: en
• Tags: transformers, safetensors, manta, feature-extraction, medical, decision-model, classification, calibration, multimodal, vision, radiology, pathology

Links

• Model: https://huggingface.co/gevaertlab/MANTA-27B
• Files: https://huggingface.co/gevaertlab/MANTA-27B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download gevaertlab/MANTA-27B

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