Nace AI has released Drex Max, a 27-billion-parameter decision model available as free open weights on Hugging Face, which classifies inputs by assigning probabilities to fixed answer sets rather than generating text. The model accepts context—including plain text, structured data, and images—alongside questions with up to 255 options, returning a scored distribution in a single forward pass; for example, it routed a duplicate-charge complaint to billing with 99.88% confidence. A 4-bit quantized version (14.45 GB) runs locally on Apple Silicon Macs (M4/M5) with 24 GB+ unified memory, while an 8-bit variant requires 48 GB; the Mac builds currently support text-only input. Nace reports the quantized version's scores deviated by ≤0.0329 from the full model on an M5 Pro with 48 GB, with identical final selections. On the Artificial Analysis Decision Index, Drex Max scores 64, ranking fourth among 18 models and outperforming Claude Haiku 5.5 (Low) at 61; on Nace's internal Decision Index 0.4 it scores 67.52 versus 61.25 for Drex v2.0. Licensed under Open RAIL-M, the model targets high-volume, low-latency routing tasks—ticket triage, content flagging, tool selection—to reduce cloud API costs and keep data on-device, though independent benchmarks on the Mac builds are not yet published.