Which MedErtas model to pick
Table with columns: If your product mostly does, Pick| If your product mostly does | Pick |
|---|
| Exam-style answers, ICD-10 coding, checking answers for hallucinations | This model: MedQA 93.0, MedConceptsQA hard 69.1, MedHallu hard 64.1 |
| Clinical calculation: doses, scores, risk calculators | MedErtas-v0.1-27B-LS: MedCalc-Bench 77.3, against 63.9 here |
| Facts from long patient records | Either: LongHealth 88.5 to 90.2 across both |
| Your own product's tasks, on your own data | A MedErtas model trained for your use case: talk to us |
Build at a glance
Built on Google's Gemma 4 31B. We quantised its weights with calibration on medical data and left the architecture unchanged.
Table | |
|---|
| Base | google/gemma-4-31B-it, architecture unchanged |
| Profile | LS v0.1: domain-aware quantisation, calibrated on medical data |
| Format | NVFP4 (W4A4) with attention in FP8. This repository holds the NVFP4 build |
| Size | 24.2 GB |
| Drafter | MedErtas-v0.1-31B-LS-Perigee, 1.0 GB: Google's Gemma 4 assistant, fine-tuned on medical answers |
| Hardware | NVIDIA Blackwell GPUs, which NVFP4 W4A4 needs. Measured on one RTX PRO 6000 (96 GB); smaller Blackwell cards are unmeasured |
| Runtime | vLLM 0.32.0 |
| Context | Measured at 65,536 tokens. The base model's longer context is inherited and unmeasured with these weights |
| Licence | Apache 2.0, from Gemma 4 |
Precision layout
Table with columns: Part, Format| Part | Format |
|---|
| MLP projections | NVFP4 |
| Attention projections | FP8 |
| Embeddings and LM head | BF16 |
| Vision and audio towers | BF16, unused when serving text |
Keeping attention in FP8 adds 0.9 GB over an all-NVFP4 layout, and it is what holds clinical calculation. The existing NVFP4 builds of Gemma 4 31B from Red Hat and NVIDIA lose 4.1 to 6.4 points on MedCalc-Bench; this build is within 1.4 points of BF16.
Results
Scored on the Medmarks harness with thinking off, on vLLM 0.32.0. Each build is paired against BF16 on the same items. Differences are in accuracy points with 95% intervals, computed per item before rounding.
Table with columns: Set (items), What it tests, BF16, This model, Δ (95% interval), Red Hat NVFP4, NVIDIA NVFP4| Set (items) | What it tests | BF16 | This model | Δ (95% interval) | Red Hat NVFP4 | NVIDIA NVFP4 |
|---|
| MedCalc-Bench 1.2 (1,100) | Clinical calculation | 65.3 | 63.9 | −1.4 (−3.7 to +1.0) | 58.9 | 61.2 |
| MedCalc-Bench 1.2, with tools (1,100) | Clinical calculation with a calculator tool | 63.8 | 61.1 | −2.7 (−5.1 to −0.4) | 59.0 | 61.7 |
| MedQA (1,273) | Exam questions | 93.2 | 93.0 | −0.2 (−1.3 to +0.9) | 93.1 | 92.5 |
| MetaMedQA (1,373) | Recognising unanswerable questions | 70.4 | 70.4 | 0.0 (−1.4 to +1.4) | 69.7 | 69.5 |
| MedHallu easy (548) | Hallucination detection | 83.6 | 85.0 | +1.5 (−0.4 to +3.3) | 83.2 | 83.9 |
| MedHallu medium (636) | Hallucination detection | 80.2 | 81.1 | +0.9 (−1.2 to +3.1) | 78.5 | 80.8 |
| MedHallu hard (816) | Hallucination detection | 64.1 | 64.1 | 0.0 (−2.0 to +2.0) | 63.5 | 62.6 |
| MedConceptsQA easy (2,000) | ICD-10 coding | 98.8 | 98.5 | −0.3 (−0.8 to +0.1) | 98.7 | 98.7 |
| MedConceptsQA medium (2,000) | ICD-10 coding | 71.7 | 70.5 | −1.1 (−2.6 to +0.4) | 70.0 | 70.4 |
| MedConceptsQA hard (2,000) | ICD-10 coding | 69.2 | 69.1 | −0.1 (−1.7 to +1.6) | 66.7 | 68.0 |
| LongHealth task 1 (400) | Facts from long patient records | 90.2 | 90.2 | 0.0 (−1.2 to +1.2) | 90.8 | 90.8 |
| LongHealth task 2 (800) | Facts from long patient records | 89.8 | 89.8 | 0.0 (−0.8 to +0.8) | 89.9 | 90.1 |
| Mean, 12 sets | | 78.4 | 78.1 | | 76.8 | 77.5 |
No set is significantly below BF16 after Holm correction across every set and build compared. MedCalc-Bench with tools is the one set whose interval falls below zero before correction. The harness samples at temperature 1.0, and each build ran once.
Speed
Single-user decode speed on clinical calculation, in output tokens per second, measured on one RTX PRO 6000 with vLLM 0.32.0 in October 2026.
Table with columns: Configuration, Decode speed (tokens/s)| Configuration | Decode speed (tokens/s) |
|---|
| Gemma 4 31B, BF16 | 22.5 |
| This model | 51 |
| This model with Google's Gemma 4 assistant | 168 |
| This model with Perigee, the MedErtas medical drafter | 189 |
With speculative decoding, the main model checks every drafted token and keeps only those it would produce itself, so answers follow this model's own distribution. The drafter changes how fast they arrive.
Serving
Text only, with Perigee. These are the settings the speed and quality results above were measured with.
vllm serve ErtasAI/MedErtas-v0.1-31B-LS \ --language-model-only \ --max-model-len 65536 \ --kv-cache-dtype fp8 \ --enable-prefix-caching \ --reasoning-parser gemma4 \ --enable-auto-tool-choice --tool-call-parser gemma4 \ --speculative-config '{"method":"mtp","model":"ErtasAI/MedErtas-v0.1-31B-LS-Perigee","num_speculative_tokens":8}'
Without Perigee, leave out --speculative-config.
Sampling, as on the Gemma 4 card and in our runs: temperature=1.0, top_p=0.95, top_k=64. Thinking is set per request with chat_template_kwargs: {"enable_thinking": false} or true; the results above are thinking off.
What makes it medical
The main weights are Gemma 4 31B's, compressed with calibration on medical data, so the compression costs as little as possible on medical tasks. Perigee, the drafter, is trained on medical answers, and the release is scored on medical evaluation sets. The main weights are not post-trained on medical data; a post-trained MedErtas build would carry the QS profile.
Train a stronger model on your data
MedErtas v0.1 is a strong starting base for medical AI products. Trained on the data your product actually handles, such as your notes, coding conventions, patient conversations and clinical workflows, it can become a much stronger model for your use case, scored against your own evaluations before it goes live. We build these for medical AI companies and healthcare enterprises, and run them on infrastructure we manage, in your own cloud or on-premise.
Talk to us about a MedErtas model trained on your data.
Intended use and limitations
MedErtas is a research model for teams building medical AI products, such as clinical documentation, patient-facing assistants and decision-support tools, as a starting point for their own validation.
- It is not a medical device, is not cleared for clinical use, and must not be used for diagnosis or treatment decisions without qualified clinical oversight.
- It has been evaluated on benchmarks only. Validate it on your own data and workflow before any use with patients.
- It inherits the behaviour and limitations of Gemma 4 31B. Like its base, it is weaker at ICD-10 code lookup than at clinical reasoning, so pair it with a coding database for coding tasks.
- Image and audio input are unmeasured with these weights.
- English only in our evaluation.
The hosted MedErtas API is in beta, with a production offering coming soon.
Licence and credits
Apache 2.0, the licence of Gemma 4. Built on Gemma 4 31B by Google DeepMind and quantised with LLM Compressor. Comparison builds from Red Hat and NVIDIA. Medical evaluation sets through Medmarks (MedARC).
Citation
@techreport{ertas2026medertas, title = {Specialisation as Inference Optimisation: Domain-Aware Quantisation and Drafting for Open Medical {LLMs}}, author = {Yang, Edward Xi}, institution = {Ertas AI}, year = {2026}, month = oct, note = {MedErtas v0.1}, url = {https://www.ertas.ai/research/medertas-v01-specialisation-inference-optimisation}}