Nace AI releases Drex Max under Open RAIL-M license

Drex Max accepts text, JSON or images as context for typed questions.[2] It supports multiple-choice, yes-or-no and ordinal-score formats, and each question can contain as many as 255 options.[3] Rather than generating a text answer, the model returns a probability for every option in one forward pass.[2] Nace AI positions it for tasks including agent routing, ticket triage and guardrail checks.[4] For teams assessing the model, that makes its stated use case a defined choice among options, rather than open-ended text generation.[2][4]

The model’s context window is 262,144 tokens, while the default request limit is 16,384 tokens.[2] Those are separate limits: the listed default request limit is lower than the model’s stated context window.[2] Images pass through a native vision tower and are visible to every question in a request.[2] The model has 64 layers, comprising 48 Gated DeltaNet linear-attention layers and 16 non-causal full-attention layers.[2] The available information describes the architecture and input handling, but does not establish how the model performs on any particular organization’s data.[2]

Nace AI says Drex Max’s text backbone runs locally on Apple silicon through llama.cpp, Ollama, Kapitsa and MLX.[2] The affine Q4 checkpoint is 14.45 GB, and Nace AI lists Apple silicon Macs with at least 24 GB of memory as compatible.[2][4] Nace AI also supports CPU inference with llama.cpp and Ollama.[2] These details give teams evaluating local deployment concrete software and memory requirements to check; the evidence does not specify performance on individual machines.[2][4]

Nace AI reports a score of 67.52 on Decision Index 0.4, which it says is 6.27 points above Drex v2.0.[4] It also reports a score of 64 on the Artificial Analysis Decision Index, ranking Drex Max fourth among 18 models.[4] On vision-related measures, Nace AI reports a public vision-index score of 73.92 across 12,210 rows and a score of 97.3 on InfographicVQA.[2] Its hosted API costs $0.03 per million input tokens and does not bill output tokens.[5] MarkTechPost disclosed that its article about the release was sponsored by Nace AI.[4]