March 08, 2026 — Cambridge, MA, USA
The rapid integration of artificial intelligence in pharmaceutical research continues to reshape the global drug discovery landscape as Liquid AI and Insilico Medicine announced a strategic partnership to develop lightweight scientific foundation models designed specifically for pharmaceutical research and drug discovery workflows. The collaboration introduces LFM2-2.6B-MMAI (v0.2.1), a unified AI model capable of delivering state-of-the-art performance across multiple drug discovery tasks within a single system, addressing a growing need among pharmaceutical companies for efficient AI tools that can operate on secure on-premise infrastructure without exposing proprietary research data to external cloud platforms.
AI-Driven Platform Designed for the Drug Discovery Pipeline
The newly developed LFM2-2.6B-MMAI model represents a major step forward in the development of specialized AI systems tailored for pharmaceutical R&D. Unlike conventional AI approaches that rely on multiple task-specific models, this architecture integrates several drug discovery capabilities into a single checkpoint, enabling seamless execution of tasks across the entire discovery pipeline. By combining Liquid AI’s efficient Liquid Foundation Model (LFM) architecture with Insilico Medicine’s MMAI Gym training platform, the collaboration demonstrates how advanced AI can support pharmaceutical research while maintaining strict data privacy and operational efficiency.
The model was trained using approximately 120 billion tokens of pharmaceutical data spanning more than 200 specialized tasks, allowing it to perform complex scientific reasoning related to medicinal chemistry, molecular optimization, and pharmacological prediction. The integrated system covers the complete drug discovery loop, including property prediction, ADMET endpoint analysis, multi-parameter molecular optimization, target-aware molecular scoring, functional group reasoning, and retrosynthesis planning. This broad capability enables pharmaceutical researchers to evaluate potential drug candidates faster and more efficiently while minimizing costly experimental failures.
According to Ramin Hasani, CEO and co-founder of Liquid AI, the development demonstrates that innovative AI architecture design rather than sheer computational scale can significantly improve scientific performance, allowing smaller models to compete with significantly larger systems while operating efficiently in controlled enterprise environments. The 2.6-billion-parameter model is designed to deliver cloud-scale scientific performance while remaining deployable entirely within private pharmaceutical infrastructure, addressing security and data-ownership concerns commonly faced by global pharma companies.
Advertisement
High-Performance Results Across Pharmaceutical Benchmarks
Initial benchmarking results indicate that the LFM2-2.6B-MMAI system achieves strong performance across several critical pharmaceutical research tasks, demonstrating its potential value for real-world drug discovery environments. In property prediction testing, the model outperformed the significantly larger TxGemma-27B model across 13 of 22 pharmacokinetic and toxicology tasks, while achieving state-of-the-art results in multiple benchmark scenarios.
The system also demonstrated exceptional performance in molecular optimization, achieving success rates of up to 98.8 percent on MuMO-Instruct multi-parameter optimization benchmarks, a widely recognized metric used to evaluate the efficiency of drug candidate design strategies. Additionally, affinity prediction testing using Insilico’s internal benchmark dataset—containing approximately 2.5 million experimental measurements across 689 protein targets—showed improved correlation scores compared to several frontier AI models, including GPT-5.1, Claude Opus 4.5, and Grok-4.1.
In medicinal chemistry applications, the model exhibited strong functional group reasoning capabilities and high-quality single-step retrosynthesis predictions, both essential features for guiding chemists during the lead optimization phase of drug discovery. These capabilities allow pharmaceutical scientists to assess chemical feasibility earlier in the development process, potentially reducing the number of unsuccessful synthesis experiments and accelerating overall development timelines.
Advancing AI-Powered Pharmaceutical Research
Industry experts believe that lightweight scientific foundation models may play an increasingly critical role in accelerating pharmaceutical innovation, particularly as companies seek to compress discovery timelines while improving the accuracy of early-stage research decisions. By enabling on-premise deployment with competitive performance, the partnership between Liquid AI and Insilico Medicine provides pharmaceutical organizations with a scalable solution for integrating AI-driven molecular design, ADMET prediction, and medicinal chemistry analysis directly into internal research pipelines.
Alex Zhavoronkov, CEO of Insilico Medicine, noted that the collaboration highlights the potential of high-efficiency scientific AI systems to transform pharmaceutical R&D, making advanced computational tools more accessible to scientists while improving the overall efficiency of drug development programs. The model’s design also aligns with Insilico’s broader Pharmaceutical Superintelligence (PSI) roadmap, which aims to integrate chemical, biological, and clinical reasoning capabilities into unified AI platforms capable of supporting the full lifecycle of therapeutic development.
As pharmaceutical companies increasingly explore AI-driven drug discovery platforms, innovations such as the LFM2-2.6B-MMAI model could play a vital role in shaping the next generation of data-driven pharmaceutical research ecosystems, ultimately enabling scientists to identify promising therapeutic candidates faster and deliver innovative treatments to patients worldwide.
z
Source: Liquid AI and Insilico press release
