Med-SEAL V1 Clinical Agent
Med-SEAL V1 refers to aagdeyogipramana/Med-SEAL-V1 — the fine-tuned base model that powers the clinical reasoning capabilities of the Med-SEAL platform.
Note:
med-r1is not deployed in the current system (requires 2× NVIDIA H200 GPUs). The production system uses SEA-LION v4-32B (AI Singapore’s National LLM) as the clinical reasoning backend with the same prompt architecture and safety guards. See Med-SEAL V1 Technical Report for the full technical report.
What is Med-SEAL V1?
Field |
Value |
|---|---|
Model name |
|
Base model |
Qwen3-VL-8B-Thinking |
Parameters |
~8B |
Modality |
Vision-Language (text + image) |
Fine-tuning |
SFT on clinical reasoning tasks |
Domain |
Chronic disease management (diabetes, hypertension, hyperlipidemia) |
Languages |
English, Chinese, Malay, Tamil |
Temperature |
0.3 |
GPU requirement |
2× NVIDIA H200 (vLLM) |
Med-SEAL V1 is purely the model — not the agent system. The multi-agent architecture (A1–A6, SEA-LION Guard, smolagents Orchestrator) consumes this model as its clinical reasoning backend.
Which Agents Use It?
Agent |
How it uses Med-SEAL-V1 |
|---|---|
A2 (Clinical Reasoning) |
Calls Med-SEAL-V1 for drug interactions, lab interpretation, condition reasoning |
A5 (Insight Synthesis) |
Calls Med-SEAL-V1 for pre-visit clinical brief generation |
A1 (Companion) |
Indirectly — delegates complex questions to A2, which calls Med-SEAL-V1 |
Other agents (A3, A4, A6) use different models (MERaLiON, SEA-LION, analytics engine) and do not use Med-SEAL V1.
Current vs Target Deployment
Current (SEA-LION) |
Target (Med-SEAL V1) |
|
|---|---|---|
Model |
SEA-LION v4-32B (AI Singapore) |
|
Endpoint |
|
Self-hosted vLLM |
GPU |
None (cloud API) |
2× H200 |
Clinical fine-tuning |
❌ (general-purpose) |
✅ (chronic disease SFT) |
Multimodal |
❌ |
✅ (vision-language) |
Platform |
GCP Cloud Run |
GCP Cloud Run or GKE |
Switching requires only environment variable changes — no code changes.