Liquid AI and Insilico Launch LFM2-2.6B-MMAI: Lightweight Model for On-Premise Drug Discovery

Liquid AI and Insilico Medicine have announced a strategic partnership to launch LFM2-2.6B-MMAI, a lightweight multi-task scientific foundation model that delivers pharmaceutical superintelligence directly on private, on-premise pharmaceutical infrastructure.

The AI revolution in drug discovery has reached a critical inflection point. For years, the industry has pursued a “scale at all costs” approach. However, massive cloud-based models frequently encounter an insurmountable barrier: data security. Pharmaceutical companies cannot risk transmitting proprietary molecular data to external cloud services. With this collaboration, the two companies have solved that tension, releasing the model to the global scientific community while ensuring full data sovereignty.

LFM2-2.6B-MMAI Matching Systems 10x Its Size

Size no longer dictates intelligence in the laboratory. At just 2.6 billion parameters, LFM2-2.6B-MMAI matches or outperforms systems ten times its size. It outperformed TxGemma-27B on 13 of 22 property prediction tasks (covering pharmacokinetics and toxicology), achieving state-of-the-art results on three tasks. 

The model also produced better correlation scores than GPT-5.1, Claude Opus 4.5, and Grok-4.1 on Insilico’s internal affinity prediction benchmark involving 2.5 million experimental measurements across 689 protein targets. These results prove that efficient architecture matters more than massive scale for science.

A Scientific Reasoning Engine Trained on Chemistry and Physics

This is not just another language model predicting text strings. Built using Liquid AI’s efficient LFM architecture and Insilico’s MMAI Gym, the team created a true scientific reasoning engine. They utilized Supervised Fine-Tuning (SFT), Reasoning Fine-Tuning (RFT), and Reinforcement Learning (RL) with proprietary reward models to teach the AI actual chemistry rules and molecular physics. The model understands the physics of molecules rather than just molecular grammar. It delivers high-quality results across the entire discovery loop—from ADMET screening to complex retrosynthesis planning.

Quality data drives performance. Trained on approximately 120 billion tokens, the model leveraged Insilico’s proprietary vault of over one billion data points, including 3D poses, binding information, 100 million+ chemical reactions, and five million experimental data points grounded in medicinal chemistry. Researchers can now ingest proprietary assays entirely within local private instances—protecting intellectual property while accelerating innovation.

The Future of On-Premise Drug Discovery

“Our ultimate goal at Insilico is Pharmaceutical Superintelligence—giving scientists superhuman tools to compress discovery timelines from years to months, and bringing life-saving therapeutics to patients faster,” says Alex Zhavoronkov, CEO of Insilico Medicine.

As Liquid AI CEO Ramin Hasani noted: “With LFM2-2.6B-MMAI, we’ve shown that efficient architecture design, not just scale, is what makes foundation models practical for the sciences.”

The model serves as a single checkpoint that replaces the need for multiple specialized tools. It achieves a high success rate in multi-parameter molecular optimization on the MuMO-Instruct benchmark. The collaboration shows that on-premise deployment can deliver competitive, state-of-the-art results across a range of drug discovery tasks. This supports the development of generalist foundation models for the sciences.

Author

Steven Chung

紅杏枝頭春意鬧固然是眾人所追求的風景，但我願意說，青蘿拂行衣更能湧生感嘆。 願在海量晦澀的生醫資訊中，為您點亮一座心燈塔。鐘煌翔