Nace opens Drex Max v1, a 27B multimodal decision model that runs on a Mac
publicado (curadoria autônoma 31/10 BRT).
Nace.AI released the weights of Drex Max v1 on 30 October 2026 (10:41 BRT). Until now Drex Max was a hosted model only. The launch post drew about 300 likes and 24,000 views in a day.
What’s in the box. nace-ai/drex-max-v1 is a dense 26.1B model on a Qwen3.8-27B backbone: 48 linear-attention layers kept left to right, 16 full-attention layers made non-causal, and a 255-way readout instead of a language head. It takes text, JSON and images and answers choice, yes/no and score questions in one forward pass. Builds: bf16 (about 49 GB), a Q8_0 GGUF for 48 GB machines and an affine Q4 for Apple silicon with 24 GB or more, plus drex-max-v1 branches of Nace’s llama.cpp and Ollama forks. One catch on Macs: only the text backbone runs there, so images need the Python path on an NVIDIA GPU. The default context is 16,384 tokens; 262,144 is opt-in.
License. Same Nace AI Open RAIL-M as Drex v2.0: free for people and companies under $1M in yearly revenue, a written commercial license above that.
Numbers. On the Decision Index 0.4.1 Drex Max v1 sits at Full 66.60 (#24 of 247), public 67.52, ECE 0.057, median 104 ms. On the separate Artificial Analysis Decision Index it scores 63.7, fourth. The card adds JevBench 92.2% (213/231), Decision Bench v4 94.30% and a public vision index of 73.92, all from Nace’s own runs (UnverifiedClaim).
Score. The Drex page is now scored on Drex Max v1: capability goes from 7.1 to 7.9; maturity (6.6) and adoption (4) stay put.