HRM-Text-1B shows architecture still beats scale

This is less about a tiny model beating a bigger one and more about proving that architecture and training objective still matter when you optimize for reasoning efficiency. The release is genuinely open, with weights, code, and a reproducible pipeline, and the reported cost/performance ratio is strong on math and reasoning benchmarks. The main caveat is that the model card describes it as a pre-alignment PrefixLM checkpoint, so it is not a drop-in general-purpose chat assistant, and the benchmark claims should be read as task-specific rather than universal. The "thinks internally" framing is shorthand for the HRM design, not evidence of human-like reasoning.