Insilico Medicine has opened its Science MMAI Gym to train open source LLMs such as Qwen and Llama into pharmaceutical-grade drug discovery engines.
Insilico Medicine has launched Science MMAI Gym (AI GYM for Science), a domain-specific training environment designed to transform open-source and proprietary large language models into pharmaceutical-grade scientific engines for real-world drug discovery and development.
Positioned as a model-agnostic infrastructure rather than a standalone model release, Science MMAI Gym enables open source LLMs such as Qwen, Llama and Mistral to be systematically trained to match or outperform specialist drug-discovery systems. Insilico reports up to 10× performance improvements across chemistry, biology and clinical reasoning benchmarks, advancing its long-term vision of Pharmaceutical Superintelligence (PSI).
The launch marks the first time Insilico has opened its internal AI training infrastructure to external partners, following more than a decade of in-house AI research and a pipeline that includes 27 preclinical candidates, 10+ IND-cleared molecules, and multiple Phase I and Phase IIa clinical trials.
Science MMAI Gym addresses a known limitation of general-purpose LLMs, which often fail on mission-critical drug discovery tasks such as ADMET prediction, toxicity assessment, pharmacokinetics and retrosynthesis. Rather than relying on prompt engineering, the Gym teaches domain-specific scientific reasoning using curated datasets, supervised fine-tuning, reinforcement learning with specialised reward models, and rigorous data decontamination.
A key open-source case study shows Qwen3-14B evolving from failing roughly 70% of medicinal chemistry benchmarks to solving over 95% of tasks within two weeks, achieving state-of-the-art or near-SOTA performance across multiple ADMET and optimisation benchmarks.
Insilico positions Science MMAI Gym as a foundational layer for Chemical Superintelligence (CSI) and Biology and Clinical Superintelligence (BSI), signalling a shift in open source drug discovery from model scale to scientific curriculum and reasoning.
