Ertas AI

Sovereign AI for every product.

Ertas AI is an applied AI research lab in Melbourne. We build specialist models, each trained for one job and refined until it is as small and fast as that job allows. Run our models through one API, train open models on your own data, or work with us on a custom model for your production agents.

New: MedErtas v0.1

MedErtas is a family of open-weight language models for medical AI companies, built for fast, low-cost serving. Each model is quantised to NVFP4 with calibration on medical data, scored against its full-precision base before release, and ships with Perigee, its medical drafter for speculative decoding on vLLM. Apache 2.0.

Scores are the mean across 12 medical sets. Speeds are single-user decode on clinical calculation. Prices in US dollars, as of November 2026. Pick the 31B for exam-style answers, ICD-10 codes and answer checking, and the 27B for clinical calculation. Each model card has the full results and the vLLM serve command.

Model page · Whitepaper · Collection · Drafters: 31B Perigee, 27B Perigee

Call MedErtas through the Ertas AI Gateway

One OpenAI-compatible API, priced per token. API inference is in beta, for test data only.

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://gateway.ertas.ai/v1",
    api_key=os.environ["ERTAS_API_KEY"],
)

resp = client.chat.completions.create(
    model="ertas-ai/medertas-27b",  # or "ertas-ai/medertas-31b"
    messages=[{"role": "user", "content": "Hello!"}],
)
print(resp.choices[0].message.content)

Get a key at app.ertas.ai and follow the gateway quickstart.

Run, train, improve

• Run. Call hosted models through the Ertas AI Gateway. MedErtas is the first family on it.
• Train. Fine-tune open models on your own data on managed GPUs, from a visual workflow, and download the model, including as GGUF for llama.cpp, Ollama and LM Studio. Start free, with paid plans from US$20 a month as of November 2026.
• Improve. For production AI agents, the adaptive model endpoint trains a custom model on your agents' traffic and puts each new version live only after it passes your evals. By invitation.

Research

• MedErtas v0.1: Specialisation as inference optimisation. How domain-aware quantisation and domain-aware drafting make a medical model faster and cheaper to serve while keeping its scores. Read the whitepaper.
• QMSum meeting summarisation. A 1.2B model trained on Ertas scored 35.41 ROUGE-1 against 32.43 for the best of five hosted proprietary models (arXiv:2609.25028, August 2026). The summariser and locator models are in this org.

More at ertas.ai/research.

In this org

• MedErtas models and their Perigee drafters.
• Research releases, starting with the QMSum summariser and locator models.
• Base models for training on Ertas: the open models in the Ertas training catalogue, in BF16 and bitsandbytes 4-bit, each with its original licence and attribution.

Work with us

Building a medical AI product, or running agents on frontier APIs at volume? We train models on your own data and run them on our infrastructure, your cloud or on-premise. Talk to us.

Website · GitHub · LinkedIn · X · Reddit · Instagram