InternLM opens three structured-decision models

InternLM has opened Intern-Decision in 0.8B, 2B, and 4B variants, a multimodal model family that maps one state plus typed questions to calibrated probabilities in one forward pass. The project repository reports 90.02% average accuracy across seven suites for 4B and 44.16 ms mean latency on an RTX 4090; its 4B model card notes that training data and private calibration records are not included. For agent builders, it is a small structured-decision runtime to test beside general-purpose models—not a drop-in chat model—and the archive’s harness cost analysis shows why per-call latency still needs task-level accounting.