Tool Suggestions
Nace Drex Max Open Weights: a 27B Decision Model That Runs on a Mac
Nace AI has released Drex Max v1 as open weights under its Nace.AI Open RAIL-M license. The distinction matters, so get it right: the hosted Drex Max model dates to mid-October, with a launch post on October 19 and a hosted price of $0.06 per 1,000 decisions. What is new on November 1 is the weights — the model itself, downloadable and runnable on your own hardware.
What a decision model is
Drex Max does not generate text. You pass it a state — text, JSON or images — plus named questions, and it returns a probability for every option in one forward pass. The question types are choice, yes-or-no, and ordinal score, with up to 255 options per question. That makes it a component for the repeated judgments inside agents and backend software: routing a request, classifying a ticket, checking a policy, approving or blocking an action.
The architecture is a dense 27-billion-parameter model with 64 layers: 48 Gated DeltaNet linear-attention layers and 16 full-attention layers placed every fourth layer, non-causal. The context window is 262,144 tokens, 16,384 by default. A native vision tower makes images visible to every question, and the model card lists a qwen3_5 model type in its configuration.
Running it locally
The release ships bf16 weights at about 49 GB, a Q8_0 GGUF build at 27.25 GB, and an affine Q4 checkpoint at 14.45 GB via MLX, plus forks of llama.cpp and Ollama. Nace documents five runtimes, including Kapitsa, and all of them serve the same POST /v1/systemone API, so switching runtimes does not mean rewriting your client.
Hardware breaks down three ways. On NVIDIA GPUs, bf16 wants an H100 and 8-bit fits a 48 GB L40S. On Apple Silicon Macs, the affine Q4 checkpoint runs with 24 GB of memory or more, and the Q8_0 GGUF build needs 48 GB. CPU inference works through llama.cpp and Ollama. One current limitation: the local Mac and CPU paths are text-only for now, with no image input — image support needs the Python implementation on an NVIDIA GPU.
On quantization, Nace’s own tests report that Q8_0 on a 48 GB Apple M5 Pro matched H100 bf16 within 0.02 per probability, and the affine Q4 build stayed within 0.0329 of Q8_0 with no choice flips. Those are the company’s measurements, and they are the reason the smaller builds are credible rather than merely smaller.
Benchmarks — all Nace’s own
Every figure in this section is self-reported by Nace AI, and the MarkTechPost article covering the release is sponsored by Nace AI, so read them as the vendor’s claims, not independent results. Drex Max scores 67.52 on the public Decision Index 0.4 for text, against 61.25 for Drex v2.0 and 57.96 for Jev 1.13.0; Nace states that version 0.4 uses the same public suite as 0.3.1. On Artificial Analysis’s Decision Index it scores 64, which Nace says ranks fourth of 18. Other reported scores include JevBench 92.2%, Decision Bench v4 94.30%, InfographicVQA skill 97.3, Tools area 81.9, and a public vision index of 73.92. The results dataset covers 140,620 requests, and unanswered requests count as wrong.
The honest summary is this: Drex Max is a specialised tool, not a general chatbot, and its value is in decisions you can audit. A probability for every option, computed locally, is easier to log, threshold and test than a paragraph of generated text — and for routing, classification and guardrail checks, that is the whole point.