Nace AI launches drex max v1: a 27b decision model for local execution on apple silicon

Nace AI has open-sourced Drex Max v1, a 27B parameter decision model for local execution on Apple Silicon. The model supports multiple quantization formats, including bf16 weights, a Q8_0 GGUF file, and an affine Q4 checkpoint. It includes forks of llama.cpp and Ollama, enabling developers to implement routing, classification, and policy checks on local hardware. Drex Max processes typed questions about a given context, returning probabilities for each option in a single forward pass. It supports three question types: choice-based, yes/no, and ordinal scoring, making it suitable for agent routing, ticket triage, and guardrail checks. The model features a dense 27B backbone with 64 layers, including 48 Gated DeltaNet linear-attention layers and 16 non-causal full-attention layers. Nace AI reports that quantization maintains performance, with the Q8_0 GGUF matching H100 bf16 results on a 48 GB Apple M5 Pro. The model is available for download, with documentation and code provided. The release includes a 140,620 request results dataset, with unanswered queries counted as incorrect. Nace AI documents five runtimes, all serving the same POST /v1/systemone API. For Apple users, the affine Q4 checkpoint requires 24 GB or more, while the Q8_0 GGUF needs 48 GB. Ollama users can offload 40 layers to Metal by setting num_gpu 40, or run on CPU with num_gpu 0. The Ollama fork requires a CAPABILITY decision line in the Modelfile. Nace AI also notes that Decision Index 0.4 uses the same public suite as 0.3.1.