Skip to content
Neural Daily

ModelsNews

Drex Max v1: an open 27B decision model that runs on a Mac

🇭🇺 Olvasd el magyarul →

Nace.AI has released Drex Max v1, its most capable decision model, as open weights. The 27-billion-parameter model does not generate text. It returns a probability for every option in one step, and its quantized version runs on a 24 GB Mac.

By: Mikola Máté Patrik · · 2 min read

Hugging Face model page of nace-ai/drex-max-v1
Image: Nace.AI · Screenshot · Screenshot of the official Hugging Face model page · Source

What a decision model is

Drex Max answers typed questions about a given context. You pass a state (text, JSON or images) and named questions, and the model returns a probability for every option in a single forward pass, with nothing generated (Hugging Face). There are three question types: choice among named options, yes/no, and an ordinal score, with up to 255 options per question.

That makes it a fit for agent routing, ticket triage and guardrail checks, where a large LLM would be slow and expensive and you only need a decision.

Under the hood

  • Dense 27B backbone with 64 layers: 48 Gated DeltaNet linear-attention layers and 16 non-causal full-attention layers.
  • 262,144-token context, 16,384 by default.
  • Native vision tower, with images visible to every question.
  • License: Nace.AI Open RAIL-M.

How good it is

On Nace's own Decision Index 0.4 it scores 67.52, versus 61.25 for its predecessor Drex v2.0. On the Artificial Analysis Decision Index it scores 64, fourth of 18 models, one point behind the leaders at 65 and ahead of Claude Haiku 5.5 (Low) at 61.

Artificial Analysis Decision Index chart with Drex Max at 64 points
Nace.AI · Press kit · Chart from the official Drex Max v1 model card · Source

Source: Nace.AI. Artificial Analysis Decision Index results as shown in the model card.

Where it runs

On NVIDIA GPUs with Python in bf16 (about 49 GB) or 8-bit (about 26 GB). On Apple silicon, the affine Q4 build (14.45 GB) runs from 24 GB of memory with Kapitsa or MLX, while the Q8_0 GGUF needs 48 GB. Nace's llama.cpp and Ollama forks handle CPU, without image support. Nace reports Q8_0 within 0.02 of bf16 probabilities and no flipped choices on Q4.

If you run routing or policy checks in the cloud, this is a good reason to test locally: data stays on the machine and cost is fixed. The MarkTechPost article was sponsored by Nace, and the figures come from the company.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles