MedErtas v0.1 31B LS

MedErtas v0.1 31B LS

Serves a medical-domain language model, a compressed Gemma 4 31B, on vLLM and NVIDIA Blackwell GPUs. Built for medical AI products that need fast, low-cost answers on exam, coding and answer-checking tasks.

Made by ErtasAI 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: MedErtas v0.1 31B LS is a language model with published weights, aimed at products in medical AI. It is Google's Gemma 4 31B, compressed with calibration on medical data so it serves faster and costs less. The card says it does best on medical knowledge, ICD-10 codes, answer checking and spotting unanswerable questions.

Who it suits: It suits teams that build medical AI products and host their own models on NVIDIA Blackwell GPUs.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Healthcare software | Product teams | Serve the model for exam-style answers and ICD-10 coding | A coding helper inside a clinical tool
Health education | Study-tool developers | Answer exam-style medical questions | A practice-question assistant
Research | Evaluation researchers | Compare the compressed build with BF16 and other NVFP4 builds | Re-run the card's medical test sets

If you have no Blackwell GPU, you can skip it, because the NVFP4 format needs one.

Where to use it:

• vLLM, version 0.32.0, the runtime the card measured
• NVIDIA Blackwell GPUs: measured on one RTX PRO 6000 with 96 GB
• Perigee, a 1.0 GB drafter model that speeds up decoding
• The Ertas AI gateway and API, which the card links as hosted options

Title: Jobs it does
Use case | Who | What they get back
Exam-style answers | Product teams | MedQA score of 93.0 in the card's tests
ICD-10 coding | Coding tool builders | MedConceptsQA hard score of 69.1
Hallucination checks | Safety reviewers | MedHallu hard score of 64.1
Facts from long patient records | Clinical software teams | LongHealth task 1 score of 90.2

How to set it up:

• An NVIDIA Blackwell GPU with enough memory for 24.2 GB of weights
• vLLM 0.32.0
• Python on Windows, macOS or Linux, for downloading the files

1. Install the Hugging Face Hub tools.
2. Download the model files.
3. Serve them on vLLM as the model card describes, with Perigee if you want the faster decoding.

pip install -U huggingface_hub
hf download ErtasAI/MedErtas-v0.1-31B-LS

How to use it: Once vLLM serves the model, you send it medical questions as text. The card scored it with thinking turned off. It names two companion builds.

Title: Related MedErtas builds
Tool | What it does
MedErtas-v0.1-31B-LS-Perigee | A 1.0 GB drafter, tuned on medical answers, that speeds up decoding
MedErtas-v0.1-27B-LS | A sibling build that scores 77.3 on MedCalc-Bench against 63.9 here

hf download ErtasAI/MedErtas-v0.1-31B-LS

Example: You ask an exam-style question, such as which ICD-10 code fits a short condition description. The model answers in text. On the card's tests it scored 93.0 on MedQA, close to the 93.2 of the BF16 original.

For newcomers and experts: If you are new, this is a large language model for medical products, shrunk from 62.6 GB to 24.2 GB so it is cheaper to serve. The first small thing to try is the two download commands, then the card's table on which MedErtas model to pick.

If you know the field, the weights use NVFP4 W4A4 and keep the attention layers in FP8, a choice that costs 0.9 GB more than pure NVFP4. The card says this holds clinical calculation within 1.4 points of BF16. The trade-off is hardware and range: Blackwell only, and a context length measured at 65,536 tokens.

Start from Gemma 4 31B
Calibrate on medical data
Quantise to NVFP4, attention in FP8
Score against BF16 on medical sets
Serve on vLLM with Perigee

Good to know:

Title: MedCalc-Bench 1.2 score
BF16: 65.3 points
This model: 63.9 points
NVIDIA NVFP4: 61.2 points
Red Hat NVFP4: 58.9 points

• Score: the card puts it at 99.6% of BF16's mean across 12 medical sets.
• Speed: a single user decodes at 189 tokens/s on calculation tasks, compared with 22.5 for BF16.
• Hardware: NVFP4 W4A4 needs NVIDIA Blackwell GPUs, and smaller cards are unmeasured.
• Context: measured at 65,536 tokens, and the base model's longer context is unmeasured here.
• License: apache-2.0, inherited from Gemma 4.
• Cost to run: a Blackwell GPU, or the hosted Ertas API, whose price the card does not give.
• Data it reaches: you download files from Hugging Face, and the card names no other access.
• Not for: calculation-heavy work such as doses and risk scores, where the card points to the 27B build.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: vllm
• Parameters: 20.9B
• Weights: about 20.9 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 59,001
• Likes: 3
• License: apache-2.0
• Languages: en
• Based on: google/gemma-4-31B-it
• Tags: vllm, safetensors, gemma4, medical, clinical, nvfp4, compressed-tensors, speculative-decoding, ertas, medertas, text-generation, conversational

Links

• Model: https://huggingface.co/ErtasAI/MedErtas-v0.1-31B-LS
• Files: https://huggingface.co/ErtasAI/MedErtas-v0.1-31B-LS/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ErtasAI/MedErtas-v0.1-31B-LS

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