Insilico Medicine and Liquid AI Launch Lightweight AI Model for On-Premise Drug Discovery
Key Insights
Insilico Medicine (search) and Liquid AI (search) have developed LFM2-2.6B-MMAI (search), a 2.6 billion-parameter AI model that achieves state-of-the-art performance across multiple drug discovery tasks while running entirely on private pharmaceutical infrastructure.
The model outperformed TxGemma-27B (search), a system more than 10 times larger, on 13 of 22 pharmacokinetics and toxicology tasks, and achieved up to 98.8% success rates on molecular optimization benchmarks.
The collaboration addresses pharmaceutical companies' need to harness advanced AI capabilities without sending proprietary molecular data to external cloud services, enabling secure on-premise deployment.
Insilico Medicine (search) and Liquid AI (search) announced a strategic partnership that has produced LFM2-2.6B-MMAI (search) (v0.2.1), a lightweight artificial intelligence model designed specifically for pharmaceutical research. The collaboration addresses a critical challenge facing drug companies: accessing cutting-edge AI capabilities while maintaining data security on private infrastructure.
Breakthrough Performance with Compact Architecture
The 2.6 billion-parameter model demonstrates that efficient architecture design, rather than scale alone, drives practical AI applications in pharmaceutical sciences. Despite its compact size, LFM2-2.6B-MMAI (search) matches or outperforms systems ten times larger across the drug discovery pipeline.
"With LFM2-2.6B-MMAI (search), we've shown that efficient architecture design, not just scale, is what makes foundation models practical for the sciences," says Ramin Hasani, CEO and co-founder of Liquid AI (search). "A single 2.6B-parameter model now matches or outperforms systems ten times its size across the drug discovery pipeline, all on private infrastructure."
Comprehensive Drug Discovery Capabilities
The model covers the complete discovery loop, spanning property prediction and ADMET endpoints, multi-parameter molecular optimization, target-aware scoring with protein-pocket conditioning, functional group reasoning, and retrosynthesis planning. Training involved approximately 120 billion tokens of pharmaceutical data across over two hundred different tasks.
Performance Benchmarks
The model achieved impressive results across multiple domains:
Property Prediction: Outperformed TxGemma-27B (search), a model more than 10 times larger, on 13 of 22 tasks covering pharmacokinetics and toxicology, achieving state-of-the-art results on three tasks when compared to specialist models built for individual tasks.
Molecular Optimization: Reached success rates of up to 98.8% on industry-standard multi-parameter optimization benchmarks (MuMO-Instruct).
Affinity Prediction: On Insilico's internal benchmark featuring 2.5 million experimental measurements across 689 protein targets (search), the model produced better correlation scores than frontier models including GPT-5.1 (search), Claude Opus 4.5 (search), and Grok-4.1 (search).
Chemical Reasoning: Demonstrated strong functional group reasoning capabilities (FGBench) and high-quality single-step retrosynthesis suggestions (ChemCensor metric).
Addressing Industry Security Concerns
The partnership tackles how pharmaceutical companies can harness cutting-edge AI capabilities without sending proprietary molecules, assays, and target data to external cloud services. By combining Liquid AI (search)'s efficient LFM architecture with Insilico's MMAI Gym, a comprehensive training platform with over 1,000 pharmaceutical benchmarks, the collaboration demonstrates that on-premise deployment can deliver competitive results across the full spectrum of drug discovery tasks in a single system.
Immediate Applications
These capabilities unlock immediately useful applications for pharmaceutical companies, particularly in high-frequency ADMET screening, medicinal chemistry-facing lead optimization, and retrosynthesis feasibility assessment that prevents wasted experimental effort.
"We are pleased to collaborate with Liquid AI (search) to develop the next generation of lightweight liquid foundation models capable of performing multiple scientific tasks with state-of-the-art performance across drug discovery benchmarks," says Alex Zhavoronkov, CEO of Insilico Medicine (search). "Highly-efficient liquid science models will make it easier for more scientists to achieve their goals in order to compress discovery timelines and ultimately help patients."
MMAI Gym Training Platform
The model was developed using Insilico's MMAI Gym for Science, a domain-specific training environment designed to elevate general-purpose Large Language Models into pharmaceutical-grade engines for drug discovery and development. The platform utilizes specialized tracks for Chemical Superintelligence (CSI) and Biology/Clinical Superintelligence (BSI) to teach models domain-specific reasoning across medicinal chemistry, biology, and clinical planning.
The curriculum leverages high-quality reasoning datasets and multi-task fine-tuning to achieve up to 10x performance gains on mission-critical R&D tasks compared to baseline models. All models are evaluated against a rigorous suite of proprietary and public benchmarks meticulously cleaned to avoid data leakage between training and test sets.