Nace AI has released the open weights of Drex Max v1, which it presents as its most performant decision model. It can now run on user hardware, including on some Macs equipped with Apple Silicon chips.
- Nace AI releases Drex Max v1, a 27-billion-parameter decision model that runs locally, including on some Apple Silicon Macs.
- The model answers questions about a given context in a single pass and returns probabilities instead of generating text.
- Nace AI claims top benchmark scores, but it provides all the measurements itself.
- The weights are available under the Nace.AI Open RAIL-M license, while the hosted version costs $0.03 per 1,000 decisions.
According to the company, the released version includes bf16 weights, a Q8_0 GGUF format file, a Q4 affine checkpoint, as well as adapted versions of llama.cpp and Ollama. For developers, the benefit is being able to keep tasks like routing, classification, or rule verification local.
A decision model, not a generation model
Drex Max responds to questions asked on a given context: a state provided as text, JSON, or images, accompanied by named questions. It returns a probability for each option in a single pass, without generating anything. Questions can be multiple choice among named options, binary (yes or no), or ordinal score.
Nace AI cites agent routing, ticket triage, and guardrail checks. The model has 27 billion parameters in a dense architecture, with a context window of 262,144 tokens, defaulting to 16,384, and up to 255 options per question. Images go through a native vision tower, and a 255-output read head transforms the last hidden state into probabilities.
Performance claims by the publisher
All cited measurements are provided by Nace AI. On the public Decision Index 0.4, Drex Max scores 67.52, compared to 61.25 for Drex v2.0 and 57.96 for Jev 1.13.0. The highlighted top score is 97.3 on the InfographicVQA benchmark. According to the company, the model achieves 92.2% on JevBench (213 correct out of 231) and 94.30% on Decision Bench v4 (1010 out of 1071). A ranking on the Artificial Analysis Decision Index places it at 64 points, in fourth place out of 18, per Nace AI. Unanswered queries are counted as incorrect.
A more or less demanding local execution
Nace AI documents five execution environments, all exposing the same POST /v1/systemone interface. With Python on NVIDIA GPU, the model occupies about 49 GB in bf16 or 26 GB in 8-bit, and accepts images. On Mac, the Q4 affine version, sized at 14.45 GB, targets machines with at least 24 GB of memory, while the 27.25 GB Q8_0 GGUF requires 48 GB. Local paths on Mac and CPU currently only support text.
According to the company’s tests, quantization maintains close accuracy: on an Apple M5 Pro with 48 GB, the Q8_0 stays within 0.02 of the results obtained in bf16 on an H100 for each probability, and the Q4 stays within 0.0329 of Q8_0, with no change in choice. The weights are distributed under the Nace.AI Open RAIL-M license. The hosted version is priced at $0.03 per 1,000 decisions at launch, according to the company.
The source article specifies that it is sponsored by Nace AI.
