Nace.AI released Drex Max, a dense 27B open-weight model built for decisions, not chat. You pass context and named questions, and it returns a probability for every option in one forward pass. Nothing is generated.
Details
- A 255-way readout head over a 64-layer backbone (48 Gated DeltaNet linear-attention layers and 16 full-attention ones)
- Scores 67.52 on the public Decision Index 0.4 and 64 on the Artificial Analysis Decision Index, fourth of 18
- On Apple silicon the affine Q4 runs on Macs with 24GB or more (Kapitsa and MLX); the Q8_0 GGUF needs 48GB. There are llama.cpp and Ollama forks
- License: the Nace.AI Open RAIL-M, a modified RAIL license with use restrictions. That’s open weights, not open source
My take
A model that just picks among options with probabilities is handy for routing and classification in agents. Read the license before building on it.
God Bless, and Use the Source!
