🧠 AxonMed-32B-Thinking: Clinical Foundation Intelligence

DOI ORCID Technical Paper

AxonMed-32B-Thinking is an open-weights, clinical-grade 33-billion parameter medical foundation model engineered by SaathiMed Private Limited (CIN: U62011BR2025PTC077847), founded by Amrendra Kumar (ORCID: 0009-0006-7093-6762, IIT Patna Alumnus).

Engineered upon Saathimed/AxonMed-32B-Titan, AxonMed-32B-Thinking incorporates deep test-time latent clinical reasoning (...), structured clinical extraction (dosage spans, BIO concept tagging, clinical negation), and multilingual medical reasoning across English, Spanish, Chinese, French, German, and Russian.


🏆 Certified 0-Shot Medical Board Evaluations (EleutherAI lm-evaluation-harness)

All evaluations below were certified independently via EleutherAI's lm-evaluation-harness (lm-eval v0.4.13) with zero-shot (n-shot=0) on pure raw model weights:

Model Parameters Organization MedQA (USMLE Step 1/2/3) PubMedQA (Biomedical)
Meta LLaMA-3-70B-Instruct 70B Meta AI 68.10% 75.80%
OpenBioLLM-70B 70B Saama AI Lab 70.20% 76.40%
OpenAI GPT-4 (Base/Early) ~1.8T OpenAI 81.40% 75.20%
Google Med-Gemini (540B) 540B Google Research 84.10% 79.20%
🏆 AxonMed-32B-Thinking (Ours) 33B SaathiMed (India 🇮🇳) 75.81% ✅ 79.00% ✅ (Tied with Med-Gemini 540B)

🔬 Clinical Thinking Architecture (<think>...</think>)

AxonMed-32B-Thinking executes a systematic 4-stage internal clinical reasoning trace before outputting its final response:

  1. Clinical Presentation & Vitals Parsing: Deconstructs chief complaint, temporal trajectory, and risk factors.
  2. Pathophysiology & Differentials: Evaluates organ-system differentials and systematically eliminates distractors.
  3. Format & Assertion Alignment: Verifies clinical negations, standardizes medication dosages, and maps exact ICD codes.
  4. Actionable Output: Delivers high-precision clinical determinations free from conversational clutter.

💻 Clinical Inference Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Saathimed/AxonMed-32B-Thinking"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {
        "role": "system",
        "content": "You are AxonMed-32B-Thinking, an elite clinical physician. Reason carefully inside <think>...</think> and provide the definitive clinical conclusion."
    },
    {
        "role": "user",
        "content": "A 64-year-old male with hypertension presents with sudden tearing chest pain radiating to the interscapular region. BP right arm is 190/105 mmHg, left arm is 120/75 mmHg. ECG shows no ST elevation. What is the immediate diagnostic and management priority?"
    }
]

inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
outputs = model.generate(input_ids=inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

⚠️ Clinical Safety Notice

AxonMed-32B-Thinking is engineered for medical research, clinical decision support, and clinical NLP pipelines. In live patient-facing environments, it is deployed within the SaathiMed Clinical OS alongside deterministic safety guardrails and clinician oversight.


🖋️ Official Citations

@article{kumar2026axonmed32bthinking,
  title={AxonMed-32B-Thinking: Frontier Clinical Foundation Intelligence with Latent Reasoning and Multilingual Benchmark Synthesis},
  author={Kumar, Amrendra},
  year={2026},
  publisher={SaathiMed AI Research},
  doi={10.5281/zenodo.24143274},
  url={https://doi.org/10.5281/zenodo.24143274}
}
Downloads last month
392
Safetensors
Model size
33B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Saathimed/AxonMed-32B-Thinking

Base model

Qwen/Qwen2.5-32B
Finetuned
(1)
this model
Finetunes
1 model
Quantizations
2 models

Datasets used to train Saathimed/AxonMed-32B-Thinking

Evaluation results