Nace.AI Open-Sources Drex Max, a 27B Model That Decides, Not Writes

While most labs compete on how well their models write and chat, Nace.AI went the other way: its new model does not generate a single word. On October 30 the company open-sourced Drex Max v1, which it calls the most capable multimodal decision model, and says it runs 100% locally on M4 and M5 Apple Silicon.

A decision model works like a multiple-choice engine. You pass context as a state, in text, JSON or images, along with one or more named questions. In a single forward pass the model returns a probability for every option, across choice, yes/no and score questions, with up to 255 options. In Nace's example, a double-charge support ticket is routed to billing with 99.88% probability, alongside a refund verdict and an urgency score, with no autoregressive generation at all.

Under the hood Drex Max is a dense 27B model. Of its 64 layers, 48 use Gated DeltaNet linear attention and 16 full-attention layers are made non-causal so every token sees the whole input, topped by a 255-way readout head. The window goes up to 262,144 tokens, but the default cap is 16,384 and has to be raised explicitly. Nace reports 67.52 on its public Decision Index 0.4, 73.92 on its vision index, and 64 on the Artificial Analysis Decision Index, fourth of 18 and one point off the leaders.

Local deployment is the other selling point. An affine Q4 checkpoint runs on Macs with 24GB or more through MLX or Nace's Kapitsa runtime; the Q8_0 GGUF needs 48GB and runs on Nace's forks of llama.cpp and Ollama. One caveat: on Macs only the text backbone runs today, and image input still needs the Python path on an NVIDIA GPU.

The appeal is obvious for enterprise plumbing. Routing, moderation, tagging and deciding which tool an agent should call are judgments, and paying for open-ended generation and then parsing it is slow, costly and unpredictable. A model that returns structured probabilities with built-in confidence is a much cleaner fit for risk, support routing and agent tool selection.

"Open source" deserves an asterisk, though. Drex Max ships under a modified Open RAIL-M license: individuals, researchers and small teams under $1 million in annual revenue and under $1 million in funding can use it commercially for free, but anyone above either line needs a separate commercial license, and use restrictions apply. Open weights with tiered pricing is the fairer description. For large companies, running locally lowers the cost of a trial; production still runs through a sales conversation.