insilicomedicine/longevity-llm
Longevity-LLM (L-LLM) is a 9-billion parameter Qwen3.5-based language model developed by Insilico Medicine, specifically adapted for aging and longevity biology. It was created through continued pretraining and supervised fine-tuning on a multi-domain corpus covering clinical aging, epigenomics, transcriptomics, proteomics, and genetics. This model excels at tasks within the longevity domain, offering specialized knowledge and reasoning capabilities for biomedical research. It supports a 32,768-token context length and is designed for text-based applications.
Loading preview...
Longevity-LLM (L-LLM) Overview
L-LLM is a 9-billion parameter language model from Insilico Medicine, built upon the Qwen3.5 architecture and specialized for the domain of aging and longevity biology. It leverages a hybrid transformer design, interleaving Gated DeltaNet linear-attention layers with standard self-attention.
Key Capabilities & Training
This model was developed through a multi-stage fine-tuning process:
- Domain Adaptation: Continued pretraining and supervised fine-tuning on a comprehensive corpus including UniProt, Gene Ontology, NHANES, GEO DNA-methylation cohorts, GTEx, TCGA, Olink plasma-proteomics, OpenGenes, SynergyAge, and CellAge data.
- Reasoning Augmentation: A reasoning-augmented continuation pass was applied using prediction tasks enhanced with frontier-model chain-of-thought traces.
- Context Length: Supports a substantial context length of 32,768 tokens.
- Deployment: Optimized for local deployment using vLLM, with specific configurations provided for various GPU VRAM capacities (e.g., H100, A100, RTX 4090).
Use Cases & Considerations
L-LLM is designed for applications requiring deep expertise in aging biology. It offers a "thinking mode" for enhanced reasoning, which can be controlled per request. Users should be aware of potential limitations, such as the fabrication of biomedical entities or incorrect answers, and are advised to use a grounding system prompt and verify all outputs. The model is text-only and is released under a CC-BY-ND-4.0 license for its fine-tuning additions.