Nace AI has open-sourced Drex Max v1, its most capable decision model to date — a dense 27-billion-parameter model that scores 67.52 on the public Decision Index 0.4 and can run locally on Apple Silicon Macs.
The release ships with bf16 weights, a Q8_0 GGUF checkpoint, an affine Q4 checkpoint, and forks of llama.cpp and Ollama. For developers, that means routing, classification and policy checks can stay on a local machine instead of the cloud. (Image Credits: MarkTechPost)
What Drex Max v1 Can Do
Drex Max is a decision model, not a chatbot. You pass it a state — text, JSON or images — plus named questions, and it returns a probability for every option in one forward pass. Nothing is generated. Question types cover choices among named options, yes/no answers and ordinal scores.
That makes it a fit for agent routing, ticket triage and guardrail checks — jobs where you need a fast, cheap judgment call rather than a long generated answer.
Drex Max v1 Benchmarks and Hardware Requirements
The model card describes a dense 27B backbone with 64 layers — 48 Gated DeltaNet linear-attention layers plus 16 full-attention layers, a 262,144-token context window, and a native vision tower feeding a 255-way readout head.
On benchmarks, Drex Max v1 hits 67.52 on Decision Index 0.4 — 6.27 points ahead of Drex v2.0 (self-reported), with a best skill of 97.3 on InfographicVQA and 81.9 in the Tools area. Nace also claims 4th of 18 on the Artificial Analysis Decision Index.
It runs on NVIDIA GPUs (H100 in bf16, or an L40S 48GB in 8-bit), Apple Silicon Macs with 24GB+ RAM (affine Q4) or 48GB (Q8_0 GGUF), and plain CPU via llama.cpp and Ollama. The MLX affine Q4 checkpoint is just 14.45 GB — genuinely portable for a 27B model.
Also Read: DeepSeek Open-Sources New Coding Model With MIT License
Source: MarkTechPost — read the full story there.
