insilicomedicine/longevity-llm
Longevity-LLM by Insilico Medicine is a 9-billion parameter Qwen3.5-based language model specifically adapted for aging and longevity biology, featuring a 32,768-token context length. It was developed through continued pretraining and supervised fine-tuning on a multi-domain corpus including clinical aging, epigenomics, transcriptomics, proteomics, and genetics. This model excels at tasks related to aging prediction and biological research, offering specialized knowledge in a domain-specific context.
Loading preview...
Longevity-LLM: A Specialized Model for Aging Biology
Longevity-LLM (L-LLM) is a 9-billion parameter model based on Qwen3.5, specifically fine-tuned and adapted by Insilico Medicine for the domain of aging and longevity biology. It leverages a hybrid transformer architecture that interleaves Gated DeltaNet linear-attention layers with standard self-attention.
Key Capabilities
- Domain-Specific Expertise: Trained on a comprehensive multi-domain corpus covering clinical aging, epigenomics, transcriptomics, proteomics, and genetics.
- Advanced Fine-tuning: Developed through a three-stage process including continued pretraining, supervised fine-tuning on aging prediction tasks, and a reasoning-augmented continuation pass.
- High Context Length: Supports a context length of 32,768 tokens, enabling processing of extensive biological data and research.
- Reasoning Mode: Features an optional "thinking mode" that can improve factual accuracy and reduce fabricated entities, though it increases token consumption.
Good For
- Aging Research: Ideal for tasks requiring deep knowledge in aging biology, including analyzing epigenetic clocks, gene annotations, and protein interactions.
- Biomedical Applications: Suitable for researchers and developers working on longevity, disease mechanisms, and drug discovery related to aging.
- Factual Inquiry: When paired with a grounding system prompt, it can provide reliable information by explicitly stating when it doesn't recognize an identifier, mitigating confident fabrication.
- Single-GPU Deployment: Optimized for deployment on a single GPU with varying VRAM configurations (e.g., H100, A100, RTX 4090), making it accessible for local inference.