Nace AI has released Drex Max v1, a 27B decision model designed for local execution on Apple Silicon hardware. The model processes text, JSON, or image inputs to provide probabilities for specific options, such as yes/no answers, multiple-choice selections, or ordinal scores, without generating text.
The architecture features a 27B backbone with 64 layers, including 48 Gated DeltaNet linear-attention layers and 16 non-causal full-attention layers. It supports a 262,144-position window and utilizes a native vision tower for image processing. The model was trained on a dataset containing 140,620 requests.
Hardware requirements vary by quantization: the affine Q4 checkpoint requires at least 24 GB of RAM, while the Q8_0 GGUF version needs 48 GB. Nace AI recommends specific configurations for Ollama users, including setting num_gpu to 40 to offload layers to Metal.
Performance testing on an Apple M5 Pro with 48 GB of RAM showed that the Q8_0 GGUF version maintained accuracy within 0.02 per probability compared to H100 bf16. The affine Q4 version remained within 0.0329 of the Q8_0 performance, with no choice flips observed during testing.
