Nace AI has released the weights for its Drex Max decision model on Hugging Face, according to the company's model page1.

The release matters because Nace says Drex Max can run locally in several configurations, including Apple silicon systems, and is designed for classification-style decision workloads that return probabilities across fixed answer choices rather than generated text.

According to the Hugging Face model page1, Drex Max is a dense 27B-parameter model with a 255-way readout head and is released under the Nace.AI Open RAIL-M license. Nace says the model takes a context and one or more named questions and returns a probability for every option "in one forward pass," adding that "nothing is generated."

The same page says Drex Max answers typed questions about a given context in text, JSON and images. Nace lists a default limit of 16,384 tokens per request and says that can be raised to the model's 262,144-token window with configuration changes on the model page1.

What Nace says was released

Nace's Hugging Face listing says the package includes bf16 weights, a separate readout head, a ready-made Q8_0 GGUF file and an affine Q4 checkpoint, along with links to code and documentation on the model page1. The page states that the bf16 weights take about 49 GB, loading the backbone in 8-bit takes about 26 GB, and the Q8_0 GGUF file is 27,250,125,216 bytes.

For local deployment, Nace says the model runs with Python on an NVIDIA GPU, while Apple silicon support is split across several paths. The company says the affine Q4 checkpoint runs on Apple silicon Macs with 24 GB of memory or more through Kapitsa and MLX, while the Q8_0 GGUF requires 48 GB on Apple silicon through llama.cpp or Ollama, according to the model page1.

Nace's GitHub repository also describes Drex Max as part of its "open decision models" lineup and says the model runs locally through python, llama.cpp, Ollama, Kapitsa and MLX, though it notes that for llama.cpp and Ollama those runners serve the text backbone rather than the full multimodal path in every case, according to the company's repository page2.

Performance claims and their limits

Nace's Hugging Face page includes several benchmark claims, but they are vendor-supplied results rather than independent testing. The company says Drex Max scored 67.52 on the public Decision Index 0.4, 73.92 on a public vision index and 64 on the Artificial Analysis Decision Index, according to the model page1.

The same page compares Drex Max with earlier Nace and third-party rows on some public evaluations. Nace says Drex Max scored 67.52 on Decision Index 0.4, compared with 61.25 for Drex v2.0 and 57.96 for Jev 1.13.0 in the table it published on the model page1. Those figures describe Nace's reported runs and table construction; the page does not include independent test results.

For hardware-specific behavior, Nace says an Apple M5 Pro system with an 18-core CPU, 20-core GPU and 48 GB of unified memory ran the Q8_0 GGUF on Metal and answered an example request within 0.02 of a bf16 run on an NVIDIA H100, according to the model page1. The company also says that, through Ollama on the same Apple M5 Pro configuration, the Q8_0 GGUF with num_gpu 40 gave the same answers as llama-server on the same page1.

Those Mac results also circulated in two posts from @testingcatalog on X3 and @testingcatalog on X4, which described the release and repeated Nace's claims about Apple silicon support and Q8_0 behavior. Those posts are not independent confirmation.

Automatica reported in October that social posts about another model release from the same X account were not confirmed by primary issuer materials in the supplied record5. In this case, Nace's Hugging Face page provides the issuer document for the Drex Max release.

Nace's GitHub repository says "Drex Max is out" and links to the model page, weights, Q8_0 GGUF and performance materials on its repository page2. The Hugging Face listing does not include pricing.