LFM2-2.6B-MMAI

AI model · Science / biology model · Text → Text by Liquid AI

Lightweight Scientific Foundation Model for Drug Discovery

A 2.6B-parameter scientific foundation model for drug discovery that runs on private infrastructure. It delivers state-of-the-art performance across property prediction, molecular optimization, and retrosynthesis while keeping proprietary data secure.

Model specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Launched
2026-03-03
Languages
English

Category: Drug Discovery

Access
On-premise deployment, private pharmaceutical infrastructure
Released
2026-03-03
Tasks Covered
200+ pharmaceutical tasks
Deployment Mode
On-premise / Private Infrastructure
Optimization Benchmark Success Rate
98.8% (MuMO-Instruct)
Architecture Type
Liquid Foundation Model (LFM)
  • 2.6B parameter scientific foundation model
  • Trained on 120B tokens of pharmaceutical data
  • 98.8% success rate on MuMO-Instruct benchmarks
  • Runs entirely on private pharmaceutical infrastructure

Overview

LFM2-2.6B-MMAI is a specialized scientific foundation model developed through a strategic partnership between Liquid AI and Insilico Medicine. Designed for pharmaceutical companies and biotech researchers, it serves as a lightweight generalist model for the entire drug discovery loop. The model operates entirely on private pharmaceutical infrastructure, allowing organizations to ingest proprietary molecules and assay data without cloud-based risks. It was trained on approximately 120 billion tokens of pharmaceutical data across more than 200 distinct tasks.

Drug Discovery Pipeline

  • Property Prediction & ADMET: The model covers property prediction and ADMET endpoints for pharmacokinetics and toxicology.
  • Retrosynthesis Planning: The model provides high-quality single-step retrosynthesis suggestions and feasibility assessments.
  • Affinity & Chemical Reasoning: It performs target-aware scoring with protein-pocket conditioning and functional group reasoning.

Enterprise Deployment

  • On-Premise Deployment Deployment: The model is designed to run entirely on private pharmaceutical infrastructure to protect proprietary molecules and assay data.
  • Lightweight Architecture: It achieves state-of-the-art results across drug discovery benchmarks with only 2.6 billion parameters.
  • Unified Generalist Model: A single checkpoint performs across multiple subdomains rather than using a patchwork of separate point models.

Benchmarks

  • MuMO-Instruct: 98.8% success rate
  • Property Prediction: Outperformed TxGemma-27B on 13 of 22 tasks

Use cases

  • Pharmaceutical researchers perform high-frequency ADMET screening and property prediction on private infrastructure.
  • Medicinal chemists use the model for multi-parameter molecular optimization to reach success rates up to 98.8%.
  • Drug discovery teams conduct retrosynthesis feasibility assessments to prevent wasted experimental effort.
  • Scientists ingest proprietary molecules and target data locally to maintain bank-grade privacy during research.

Specifications

Domains
Pharmaceutical research, drug discovery, medicinal chemistry, biology
Access
On-premise deployment, private pharmaceutical infrastructure
Released
2026-03-03
Tasks Covered
200+ pharmaceutical tasks
Deployment Mode
On-premise / Private Infrastructure
Optimization Benchmark Success Rate
98.8% (MuMO-Instruct)
Architecture Type
Liquid Foundation Model (LFM)

How it works

  1. Collaborative Development: Liquid AI and Insilico Medicine co-designed the model by combining Liquid's LFM architecture with Insilico's MMAI Gym training platform.
  2. Domain-Specific Training: The model is trained on 120 billion tokens across 200+ pharmaceutical tasks including property prediction and molecular optimization.
  3. Private Deployment: Enterprises deploy the single 2.6B-parameter checkpoint on their own private infrastructure to ingest proprietary data locally.

FAQ

How does the model handle data privacy?

LFM2-2.6B-MMAI is designed for on-premise deployment within local private instances, ensuring that sensitive PII, proprietary molecules, and transaction data remain locked in legacy and on-prem cores.

What data was used to train the model?

The model was trained on approximately 120 billion tokens of pharmaceutical data covering over two hundred different tasks.

What specific drug discovery tasks can it perform?

It covers the complete discovery loop, including property prediction, ADMET endpoints, multi-parameter molecular optimization, target-aware scoring, functional group reasoning, and retrosynthesis planning.

Added by AntSpace from public sources (liquid.ai).

Related

More AI models by Liquid AI