Language models are finally competitive with purpose-built software in narrow chemistry jobs, Insilico Medicine argues. It released a family of compact specialist models on September 1 to back the claim.
The tools come from an internal training engine called MMAI Gym. Rather than pushing one broad model to handle everything, the company tunes small language models on clusters of related scientific tasks. Early results, it says, span more than 70 benchmarks in chemistry, biology and aging research.
The chemistry set covers synthesis planning, safety and metabolism prediction, and potency scoring across large receptor panels. Models built for target activity were tuned on 44 GPCRs and 67 kinases, and Insilico says their potency forecasts set new marks. The areas involved, such as drug interactions and toxicity, are exactly where dedicated software used to hold the edge.
Chief executive Alex Zhavoronkov argues the results reset expectations for what small models can do. He described the release as a step toward predictive AI in drug research, where domain-specific training replaces general-purpose capability as the goal. The launch extends Insilico’s push to standardize how AI models are scored for drug discovery.
