Nace.AI's Drex Max Skips Token Decoding to Answer 255-Way Decisions Instantly
Nace.AI open-sourced Drex Max v1, a 27B multimodal decision model that returns calibrated probabilities instead of generated text, running locally on Apple Silicon.
Takeaways
- Nace.AI open-sourced Drex Max v1, a dense 27B multimodal decision model under Open RAIL-M.
- Returns calibrated probabilities over up to 255 options in one forward pass, no token generation.
- Scores 67.52 on public Decision Index 0.4 and 64 on Artificial Analysis Decision Index.
- Hybrid backbone: 48 Gated DeltaNet linear-attention layers plus 16 non-causal full-attention layers, 262K context.
- Runs on NVIDIA (bf16/8-bit), plus Apple Silicon via llama.cpp, Ollama, Kapitsa, MLX.
- Q8_0 GGUF needs 48 GB; affine Q4 checkpoint runs on 24 GB Macs.
Nace.AI has released the Drex Max model card and weights for a 27-billion-parameter model built for typed decision tasks. Given context plus one or more multiple-choice, yes-or-no, or rating-scale questions, it returns a probability for every option in a single forward pass. Its fixed readout removes token-by-token decoding and output parsing from classification workflows.
The release includes bf16, 8-bit, GGUF, and affine Q4 variants under the Nace.AI Open RAIL-M license. The Q8_0 GGUF runs on Apple silicon via llama.cpp, Ollama, Kapitsa and MLX, and the affine Q4 checkpoint runs on Apple silicon Macs with 24 GB of memory or more.
The 255-way decision head
Most open-weight models expose a language-modeling head that predicts the next token. Drex Max uses a 255-way classification head that maps its final hidden representation to as many as 255 option slots. The runtime groups the relevant scores by question and normalizes them into separate probability distributions.
Questions use three schemas:
- Choice: Selects among named options, each with an optional natural-language criterion.
noul: Returns probabilities for a yes-or-no decision.- Rating: Produces a continuous score and probabilities for the discrete scale bins.
Because inference ends after one network evaluation, latency excludes the autoregressive decoding used by chat models. A request containing three questions returns three distributions without generating explanatory text.
Full-prompt attention changes the runtime
The backbone has 64 layers: 48 Gated DeltaNet linear-attention layers that stay left to right, and 16 full-attention layers (every fourth) that are non-causal. The DeltaNet layers process preceding tokens efficiently, while the full-attention layers connect every prompt position. Images pass through the native vision tower and remain available to every question in the request.
The non-causal layers require the bundled loader and the classification weights in readout.safetensors. Running the raw backbone through a standard causal chat path changes the computation and degrades the reported scores. Compatible deployments therefore need both the custom attention behavior and the readout head.