Overview
AnewMind is an internal LLM at the hundred-billion-parameter scale, integrated into AnewDDE. It is enhanced through full-parameter post-training for pharmaceutical knowledge, structure-property reasoning, and multi-stage R&D decision-making. It supports ADMET and developability analysis, relative ranking, candidate prioritization, and complex drug discovery reasoning. Its capabilities are systematically evaluated across ADMET and developability tasks, as well as on PharmBench, an internal benchmark designed by pharmaceutical R&D experts. Evaluation results show that AnewMind is competitive with leading frontier models worldwide in both ADMET analysis and pharmaceutical R&D reasoning.
Key Results
Balancing Domain Specialization with General Capabilities
After full-parameter post-training in the pharmaceutical domain, AnewMind remains competitive on benchmarks of general knowledge and scientific reasoning, including MMLU-Pro and GPQA Diamond, indicating that its enhanced pharmaceutical R&D capabilities are achieved while largely preserving broad knowledge and reasoning abilities.
Figure 1. General benchmark performance on MMLU-Pro (left) and GPQA Diamond (right). Bars show accuracy, with AnewMind Preview highlighted in blue and external models shown in gray.
Cross-modality ADMET and Developability Evaluation
AnewMind evaluates candidates through relative ranking aligned with R&D prioritization rather than isolated point estimates, covering small-molecule ADMET, cyclic-peptide permeability, and antibody developability. In standardized comparisons with frontier models, AnewMind ranked among the top five in all four aggregate evaluations, placed first in cyclic-peptide permeability, achieved near-leading performance in antibody developability, and obtained leading results across multiple CYP inhibition endpoints.
Figure 2. ADMET and developability benchmark. Prediction performance across small-molecule ADMET classification and regression, peptide permeability, and antibody developability. Colored bars highlight AnewMind Preview; higher scores indicate better performance.
PharmBench for Long-Horizon Reasoning in Pharmaceutical R&D
PharmBench was designed by pharmaceutical R&D experts around real-world drug discovery cases. By progressively revealing project context and experimental evidence, it evaluates evidence integration and R&D decision-making. AnewMind demonstrated consistent strengths across most capability dimensions and cases, with overall performance competitive with leading frontier models worldwide.
Figure 3. Construction and capability coverage of PharmBench. (a) Experts integrate experimental evidence and R&D context to formulate scenario-based questions. (b) The 200 subquestions span seven capability dimensions.
Figure 4. Overall and capability-level performance on PharmBench. Bar charts show weighted scores across 17 models, with AnewMind Preview highlighted in color.