MedErtas v0.1 27B LS

MedErtas v0.1 27B LS

Open-weight medical language model squeezed to 22 GB, best at clinical calculations and facts from long patient records. For teams building medical AI products on NVIDIA Blackwell GPUs.

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 27B LS is a medical language model from ErtasAI, built on Qwen3.8-27B. The team quantized its weights using medical data for calibration and left the architecture alone. The result is a 22 GB build, down from 55.6 GB at BF16, meant for fast, low-cost serving.

Who it suits: Teams building medical AI products that need clinical calculations or answers drawn from long patient records. The card says it is strongest on those two jobs. Treat its answers as drafts for a qualified person to check.

Title: Who uses it, and how
Industry | Who uses it | How they use it | Example
Health software | Product teams | Power calculators for doses, scores and risk | A tool computes a clinical score from entered patient values
Clinical data | Data and records teams | Pull facts out of long patient records and return short answers | A reviewer asks what a record says about one test result
Medical AI research | Evaluators | Compare a compressed build with its full-precision base | Rerun the card's medical test sets on your own setup
Infrastructure | ML platform engineers | Serve a 27B medical model at lower cost on one GPU | Run it on vLLM on a Blackwell card

Skip it if your product mostly does exam-style answers, ICD-10 coding or hallucination checks, because the card points to the 31B-LS sibling for those.

Where to use it:

• vLLM on NVIDIA Blackwell GPUs, which the 4-bit format needs
• The Ertas AI gateway and API, which the card offers as a hosted route
• Windows, macOS or Linux for downloading the files

Use case | Who | What they get back
Clinical calculation | Product teams | Doses, scores and risk numbers, scored 77.3 on the card's calculation set
Long records | Data teams | Short answers drawn from long patient records, scored at 88.5 to 89.4
Fast serving | Platform teams | Quick decoding, using the Perigee drafter the card names

How to set it up: You need Python and pip to download, plus an NVIDIA Blackwell GPU to run it. The card measured one RTX PRO 6000 with 96 GB and says smaller Blackwell cards are unmeasured.

• Python and pip
• An NVIDIA Blackwell GPU, since the NVFP4 format needs one
• vLLM at the version the card lists, 0.32.0
• Optionally, the Perigee drafter of about 4.0 GB

1. Install the Hugging Face hub library.
2. Download this repository.
3. Serve it with vLLM on a Blackwell GPU.

pip install -U huggingface_hub
hf download ErtasAI/MedErtas-v0.1-27B-LS

How to use it: Serve the model with vLLM and send it clinical questions or calculation tasks through the server. The card tested with thinking turned off. Add the Perigee drafter for faster decoding.

Tool | What it does
vLLM | Serves the model, tested at version 0.32.0
Perigee drafter | A 4.0 GB helper model that speeds up decoding on medical text
Ertas AI gateway | A hosted way to use the model without your own GPU

Example: You give the served model a patient's lab values and ask it to compute a risk score. It works through the calculation and returns a figure. A person should check the result, since the card measured 77.3 on its calculation set.

For newcomers and experts: If you are new, this is a medical specialist model shrunk to fit one high-end GPU. It needs newer NVIDIA hardware, so start by checking you have a Blackwell card.

For experts, the build uses NVFP4 with W4A4 for the MLP projections and LM head, FP8 for attention, and BF16 for gating, embeddings and the vision tower. That matches NVIDIA's NVFP4 layout, and the difference is medical calibration. The card reports 98.2 percent of BF16's mean score across 12 sets and 207 tokens per second on clinical calculation, against 26 at BF16.

Start from Qwen3.8-27B
Calibrate on medical data
Quantize to NVFP4
Serve on vLLM with a Blackwell GPU
Pair with the Perigee drafter

Good to know:

Title: Mean score across 12 medical sets
BF16: 78.4
This model: 77
Red Hat NVFP4: 76.4
NVIDIA NVFP4: 76.7

• License: Apache 2.0, from Qwen3.8
• Cost to run: none named for local use beyond a Blackwell GPU, and the hosted Ertas API is a separate service
• Data it reaches: it downloads files from Hugging Face, and a hosted call would send your prompts to Ertas
• Not for: GPUs older than Blackwell, or contexts beyond the 65,536 tokens the card measured
• Some scores fall a few points below BF16, such as 3.6 on clinical calculation

On Hugging Face, 13 November 2026

• Task: text generation
• Library: vllm
• Parameters: 18.2B
• Weights: about 18.2 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 4,187
• Likes: 6
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.8-27B
• Tags: vllm, safetensors, qwen3_5, medical, clinical, nvfp4, modelopt, speculative-decoding, qwen3.8, ertas, medertas, text-generation

Links

• Model: https://huggingface.co/ErtasAI/MedErtas-v0.1-27B-LS
• Files: https://huggingface.co/ErtasAI/MedErtas-v0.1-27B-LS/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ErtasAI/MedErtas-v0.1-27B-LS

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