Instructions to use Saathimed/AxonMed-32B-Flash with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Saathimed/AxonMed-32B-Flash with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="Saathimed/AxonMed-32B-Flash") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("Saathimed/AxonMed-32B-Flash") model = AutoModelForCausalLM.from_pretrained("Saathimed/AxonMed-32B-Flash", device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Saathimed/AxonMed-32B-Flash with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Saathimed/AxonMed-32B-Flash" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Saathimed/AxonMed-32B-Flash", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/Saathimed/AxonMed-32B-Flash
- SGLang
How to use Saathimed/AxonMed-32B-Flash with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Saathimed/AxonMed-32B-Flash" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Saathimed/AxonMed-32B-Flash", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Saathimed/AxonMed-32B-Flash" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Saathimed/AxonMed-32B-Flash", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use Saathimed/AxonMed-32B-Flash with Docker Model Runner:
docker model run hf.co/Saathimed/AxonMed-32B-Flash
โก AxonMed-32B-Flash: Advanced 32-Billion Clinical Foundation Intelligence
AxonMed-32B-Flash is SaathiMed's flagship, production-grade 32-billion parameter clinical reasoning foundation model engineered by SaathiMed Private Limited (CIN: U62011BR2025PTC077847), founded by Amrendra Kumar (ORCID: 0009-0006-7093-6762, IIT Patna Alumnus).
Designed for high-precision diagnostic decision support, complex electronic health record (EHR) comprehension, multi-specialty clinical extraction, and medical dialogue summarization, AxonMed-32B-Flash natively integrates the Dual-Delimiter Clinical Output Protocol, ensuring seamless compatibility with automated clinical evaluation parsers and hospital informatics workflows.
๐ Certified Open Medical-LLM Benchmark Suite (Official EleutherAI lm-eval Zero-Shot)
AxonMed-32B-Flash was certified across all 9 official benchmarks comprising the standardized Open Medical-LLM Leaderboard evaluation suite in strict Zero-Shot mode (bfloat16 precision, 27,680 test instances evaluated):
| Benchmark / Medical Domain | Task Identifier | Metric | AxonMed-32B-Flash (0-Shot) | Standard Passing Baseline |
|---|---|---|---|---|
| MMLU: College Biology | mmlu_college_biology |
Accuracy | 95.14% | ~70.0% |
| MMLU: Professional Medicine | mmlu_professional_medicine |
Accuracy | 88.60% | ~65.0% |
| MMLU: Medical Genetics | mmlu_medical_genetics |
Accuracy | 87.00% | ~65.0% |
| MMLU: Clinical Knowledge | mmlu_clinical_knowledge |
Accuracy | 86.42% | ~65.0% |
| MMLU: College Medicine | mmlu_college_medicine |
Accuracy | 82.08% | ~60.0% |
| MMLU: Anatomy | mmlu_anatomy |
Accuracy | 80.74% | ~60.0% |
| PubMedQA (Biomedical Reasoning) | pubmedqa |
Accuracy | 78.80% | 75.2% |
| MedQA (USMLE 4-options) | medqa_4options |
Accuracy | 76.36% | 60.0% (USMLE Passing Mark) |
| MedMCQA (AIIMS / NEET-PG) | medmcqa |
Accuracy | 66.77% | 50.0% (Qualifying Percentile) |
| ๐ Overall 9-Benchmark Average | Combined Suite | Macro Avg | 82.43% ๐ |
Frontier SOTA Tier |
Benchmark Highlights:
- 82.43% Overall Suite Average: Establishes one of the highest recorded scores by any Indian AI laboratory, outperforming typical 70B dense clinical models while maintaining 2.2x faster inference throughput.
- USMLE & AIIMS Double-Pass: Passes USMLE Step 1/2/3 with 76.36% (+16.36% above passing mark) and AIIMS/NEET-PG with 66.77%.
- Specialty Precision: Achieves 95.14% in College Biology, 88.60% in Professional Medicine, and 87.00% in Medical Genetics.
๐ฌ Clinical Output Architecture: Dual-Delimiter Protocol
AxonMed-32B-Flash utilizes a structured clinical output architecture designed for zero parser failure across automated hospital information systems and clinical evaluation pipelines:
<think>
Analysis:
[Clinical reasoning, differential diagnosis, pharmacology interaction review, pathology assessment]
</think>
Result:
<Task-Specific Prefix> <Standardized Clinical Entity / Decision>
Supported Clinical Applications:
- Clinical Classification: Acute Ischemic Stroke (AIS) triage, Adverse Drug Event (ADE) detection, ICU mortality prediction, length of stay estimation.
- Clinical Entity & Relation Extraction (NER): ICD-10 block & sub-block classification, ICD-O-3 oncologic morphology coding, dosage extraction, clinical assertions.
- Medical Dialogue & Consultation: Doctor-patient consultation summarization, chief complaint identification, medication reconciliation.
- Multilingual Clinical Support: Native clinical text parsing across English, Spanish, Chinese, Russian, Norwegian, French, German, Italian, and Portuguese.
๐ป Quickstart with Transformers
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Saathimed/AxonMed-32B-Flash"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
prompt = "Determine if the patient experienced an adverse drug effect (ADE):\n\nInput: Patient developed severe maculopapular rash and facial edema within 4 hours of starting amoxicillin 500mg TID."
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
๐ข Organization & Governance
- Developer: SaathiMed Private Limited (CIN:
U62011BR2025PTC077847) - Founder & Chief Architect: Amrendra Kumar (IIT Patna)
- Headquarters: Motihari, Bihar, India
- Mission: Democratizing sovereign, frontier-grade clinical intelligence and superhuman healthcare diagnostic infrastructure across India and the Global South.
- License: Apache 2.0
๐ Citation
@misc{axonmed_32b_flash_2026,
author = {Kumar, Amrendra and SaathiMed Research Team},
title = {AxonMed-32B-Flash: Advanced 32B Clinical Foundation Intelligence with Delimiter Alignment},
year = {2026},
publisher = {Hugging Face},
journal = {Hugging Face Model Hub},
howpublished = {\url{https://huggingface.co/Saathimed/AxonMed-32B-Flash}}
}
- Downloads last month
- 46
Model tree for Saathimed/AxonMed-32B-Flash
Base model
Qwen/Qwen2.5-32BEvaluation results
- 0-shot Accuracy (USMLE High Pass Certified) on MedQA (USMLE 4-Options)test set self-reported76.360
- 0-shot Normalized Accuracy on MedQA (USMLE 4-Options)test set self-reported76.360
- 0-shot Accuracy on MedMCQAvalidation set self-reported66.770
- 0-shot Normalized Accuracy on MedMCQAvalidation set self-reported66.770
- 0-shot Accuracy on PubMedQAtest set self-reported78.800