Nace.AI Releases Drex Max v1: Open-Weight 27B Multimodal Decision Model
Nace.AI has released Drex Max v1, an open-weight 27B multimodal decision model that utilizes a hybrid linear and full-attention architecture with a non-autoregressive readout head. Instead of generating text tokens sequentially, it processes the entire context and outputs probabilities for up to 255 candidate answers in a single forward pass. This model introduces a non-autoregressive inference approach tailored for high-speed local decision-making and tool-use tasks on consumer-grade and enterprise hardware, including Apple Silicon. By removing the traditional decoding loop, it achieves rapid response times of around 400ms while maintaining high accuracy across structured benchmarks. Drex Max v1 features a 64-layer hybrid backbone comprising 48 Gated DeltaNet linear-attention layers and 16 bidirectional full-attention layers, supporting up to 262K context length and 4-megapixel images. It is distributed under the Open RAIL-M license and runs locally via custom forks of llama.cpp and Ollama using Q8_0 GGUF quantization.
## BACKGROUND
Linear attention mechanisms like Gated DeltaNet approximate standard attention to process long sequences more efficiently by avoiding the quadratic memory scaling of traditional transformers. Non-autoregressive models bypass token-by-token generation, enabling parallel prediction of complex outputs in a single forward pass.