Nace AI's Drex Max v1 is a 27B decision model that returns probabilities instead of text

Nace AI released Drex Max v1 on 30 October. It is a dense 27B model with a 255-way readout head: instead of writing an answer, it reads text, JSON or images plus a list of options and returns a probability for each option in a single forward pass. That makes it a decision model, the same category as Baseten’s mammoth-1 and elephant-1 and Liquid AI’s d1.

Scores and hardware

• Benchmarks: 67.52 on Decision Index 0.4 and 64 on the Artificial Analysis Decision Index, which Nace’s table ranks fourth of 18. These are Nace’s own figures.
• Context: 16,384 tokens by default, raised to 262,144 with the DREX_MAX_CONTEXT setting.
• Memory: about 49 GB in bf16. A 4-bit MLX build runs on Macs with 24 GB or more, and a Q8_0 GGUF needs 48 GB. The GGUF builds use forks of llama.cpp and Ollama, not upstream releases.

The licence

Drex Max ships under the Nace.AI Open RAIL-M licence, and it is not an open-source licence. Personal and research use is allowed. Commercial use needs a separate agreement if your organization had more than $1M in gross revenue in the prior 12 months, has raised more than $1M, or offers a competing product. The use restrictions also bar, among other things, automated decisions with legal effects on people and military use.

Why this matters

Decision models fit jobs where you want one calibrated choice and a confidence number, such as routing a ticket, picking a tool or flagging a policy breach, without parsing free text. A 27B model in that class that runs on a 24 GB Mac is new. The licence decides who can use it: hobbyists and researchers can start today, while funded startups and companies need to talk to Nace before production, and anyone making decisions about people is excluded.