Model reference · open weights

Llama-3-UltraMedical

LLMs TsinghuaC3I Text gen 1 build Open, with conditions 106k dl/mo

Llama-3-UltraMedical is an open-weight language model from TsinghuaC3I. Llama-3-8B-UltraMedical (BF16) weighs 16.1 GB; the smallest configuration that runs it is 2× RTX 3060 12 GB.

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

  • Llama-3-UltraMedical is an 8.0B parameter large language model developed by TsinghuaC3I for biomedical text generation.
  • It is fine-tuned on the UltraMedical dataset to support medical examination access, literature comprehension, and clinical knowledge tasks.
  • The model supports a context length of 8192 tokens and is distributed under the llama3 license.

Summary of the TsinghuaC3I/Llama-3-8B-UltraMedical model card, 2026-10-01

What it is

Released byTsinghuaC3I
Released2024-04-27
Parameters8.0B
VRAM16.1 GB for the weights

What it runs on

Memory and cards for Llama-3-8B-UltraMedical (BF16)

16.1 GBweights, file size
131 MBcache per 1K tokens
745 MBruntime overhead, at least
8,192 tokenscontext max
CardRequests at onceContext maxMemory
8K each
RTX 3060 12 GB … RTX 4060 Ti 16 GB
2 smaller cards
RTX 3090 24 GB6all 8K23.4 GB
RTX 4090 24 GB6all 8K23.4 GB
RTX 5090 32 GB13all 8K31.0 GB
L40S 48 GB25all 8K44.0 GB
A100 80 GB57all 8K78.2 GB
H100 80 GB53all 8K78.1 GB
RTX PRO 6000 Blackwell 96 GB67all 8K93.8 GB
DGX Spark (GB10) 128 GB unified80all 8K107 GB
H200 141 GB109all 8K138 GB
B200 180 GB143all 8K176 GB
2× RTX 3060 12 GB
tensor parallel
5all 8K11.6 GB a card
2× RTX 4060 Ti 16 GB
tensor parallel
12all 8K15.4 GB a card
2× RTX 4090 24 GB
tensor parallel
27all 8K23.4 GB a card
2× RTX 3090 24 GB
tensor parallel
27all 8K23.4 GB a card
Memory needed at each load
Requests at once8K tokens each
117.9 GB
522.2 GB
825.4 GB
1634.0 GB
3251.2 GB
6485.5 GB

One card, with vLLM's small-card settings.

From the model card

What TsinghuaC3I says about Llama-3-UltraMedical

Read the model card

Experience it in our 🤗 Huggingface Space Demo!

Llama-3-8B-UltraMedical is an open-access large language model (LLM) specialized in biomedicine. Developed by the Tsinghua C3I Lab, this model aims to enhance medical examination access, literature comprehension, and clinical knowledge.

Building on the foundation of Meta's Llama-3-8B, Llama-3-8B-UltraMedical is trained on our UltraMedical dataset, which includes 410,000 diverse entries comprising both synthetic and manually curated samples.

Llama-3-8B-UltraMedical has achieved top average scores across several popular medical benchmarks, including MedQA, MedMCQA, PubMedQA, and MMLU-Medical. In these benchmarks, Llama-3-8B-UltraMedical significantly outperforms Flan-PaLM, OpenBioLM-8B, Gemini-1.0, GPT-3.5, and Meditron-70b. We extend our gratitude to Meta for the Llama model, which provided an excellent foundation for our fine-tuning efforts.

Usage

Input Examples

This model utilizes the Llama-3 default chat template without a system prompt. Below, we provide input examples for multi-choice QA, PubMedQA, and open-ended questions.

Note: To reproduce our evaluation results for the medical QA benchmark, we recommend using the following format to organize questions and multiple-choice options.

  • Input example for MedQA and MedMCQA:
A 42-year-old homeless man is brought to the emergency room after he was found unconscious in a park. He has alcohol on his breath and is known to have a history of chronic alcoholism. A noncontrast CT scan of the head is normal. The patient is treated for acute alcohol intoxication and admitted to the hospital. The next day, the patient demands to be released. His vital signs are a pulse 120/min, a respiratory rate 22/min, and blood pressure 136/88 mm Hg. On physical examination, the patient is confused, agitated, and sweating profusely, particularly from his palms. Generalized pallor is present. What is the mechanism of action of the drug recommended to treat this patient_s most likely condition?

A. It increases the duration of GABA-gated chloride channel opening.
B. It increases the frequency of GABA-gated chloride channel opening.
C. It decreases the frequency of GABA-gated chloride channel opening.
D. It decreases the duration of GABA-gated chloride channel opening.
  • Input example for PubMedQA: We organize the context and questions in a multi-choice format, similar to MedPrompt.
Context: Pediatric glioblastoma is a malignant disease with an extremely poor clinical outcome. Patients usually suffer from resistance to radiation therapy, so targeted drug treatment may be a new possibility for glioblastoma therapy. Survivin is also overexpressed in glioblastoma. YM155, a novel small-molecule survivin inhibitor, has not been examined for its use in glioblastoma therapy.
Context: The human glioblastoma cell line M059K, which expresses normal DNA-dependent protein kinase (DNA-PK) activity and is radiation-resistant, and M059J, which is deficient in DNA-PK activity and radiation-sensitive, were used in the study. Cell viability, DNA fragmentation, and the expression of survivin and securin following YM155 treatment were examined using MTT (methylthiazolyldiphenyl-tetrazolium) assay, ELISA assay, and Western blot analysis, respectively.
Context: YM155 caused a concentration-dependent cytotoxic effect, inhibiting the cell viability of both M059K and M059J cells by 70% after 48 hours of treatment with 50 nM YM155. The half-maximal inhibitory concentration (IC50) was around 30-35 nM for both cell lines. Apoptosis was determined to have occurred in both cell lines because immunoreactive signals from the DNA fragments in the cytoplasm were increased 24 hours after treatment with 30 nM YM155. The expression of survivin and securin in the M059K cells was greater than that measured in the M059J cells. Treatment with 30 nM YM155, for both 24 and 48 hours, significantly suppressed the expression of survivin and securin in both cell lines.
Does novel survivin inhibitor YM155 elicit cytotoxicity in glioblastoma cell lines with normal or deficiency DNA-dependent protein kinase activity?

A. maybe
B. yes
C. no
  • Input example for open-ended questions:
hi doctor,i am chaitanya.age 28,from hyderabad.my problem is ....i got thyroid in my frist preganacy .my delivary date was on july 24th 2009 but on july 6th early morning around 7 oclock suddenly heany bleeding started and i rushed to the hospital but they could not save the baby(boy)...i lost my frist baby.then after 6 month i concevied again but doctors said that baby is having some heart problem and the sevarity of the problem can be known after the baby birth and i should go for a planned delivery.doctors did a c section on cotober 21 2010.doctors said that babys problem is not that serious but it is a heart problem so we need wait and see for 7 days.on 5th day the baby is dead.i want to know is their any problem in me that it is happing like this...do i need o go for any test before planning for next baby.i had 2 c section till now.what are the chances for me for the next baby.how long do i need to wait and plan for next preganacy.
Investigate the mechanistic implications of statins, primarily used for lipid modulation, on the immunomodulatory pathways, with an emphasis on delineating their therapeutic impact in the context of managing clinical outcomes for individuals afflicted with cardiovascular diseases, including a requirement to discuss the implications for atherosclerotic disease progression.

Inference with vLLM

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

llm = LLM(model="TsinghuaC3I/Llama-3-8B-UltraMedical", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("TsinghuaC3I/Llama-3

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms