Nace AI released Drex Max v1 as open weights on November 1, 2026 — its most capable decision model, now downloadable and runnable on your own hardware instead of only through its hosted API. Why it matters: routing, classification and policy checks are the quiet decisions AI agents make between steps, and this release lets developers run that layer locally on an Apple Silicon Mac or an NVIDIA GPU, with no data leaving the machine.
November 1, 2026 — AI News
What happened
Drex Max is not a chatbot. It generates no text at all. You send it a state — text, JSON or images — plus typed questions, and it returns a probability for every option in a single forward pass. Questions come in three types: a choice among named options, a yes/no, and an ordinal score. That makes it a fit for agent routing, ticket triage and guardrail checks, where you need a scored answer rather than a paragraph.
The open release, Drex Max v1, is a dense 27-billion-parameter model with a 262,144-token window (16,384 by default) and support for up to 255 options per question. It ships in several formats — bf16 weights, a Q8_0 GGUF build and an affine Q4 checkpoint — along with forks of llama.cpp and Ollama, under Nace AI’s Open RAIL-M license.
The details
According to the model card reported by MarkTechPost, Drex Max uses a 64-layer backbone: 48 Gated DeltaNet linear-attention layers plus 16 full-attention layers placed every fourth layer. Images pass through a native vision tower, and a 255-way readout head turns the final hidden state into option probabilities.
On hardware, the Python runtime runs on NVIDIA GPUs (an H100 in bf16, or a 48 GB L40S in 8-bit). On Apple Silicon, the affine Q4 checkpoint targets Macs with 24 GB or more of memory, while the Q8_0 GGUF needs 48 GB. CPU paths run through the llama.cpp and Ollama forks — though the local Mac and CPU paths are text-only for now. Nace says its Q8_0 build on a 48 GB Apple M5 Pro matched H100 bf16 output within 0.02 per probability in its own tests.
How Drex Max compares
| Model | Parameters | Weights | Public Decision Index |
|---|---|---|---|
| Drex Max v1 | 27B dense | Open (Open RAIL-M) | 67.52 (v0.4) |
| Drex v2.0 | 35B MoE, ~3B active | Open (Open RAIL-M) | 61.25 (v0.3.1) |
| Drex v1.5 | 9B | Open (Open RAIL-M) | 58.08 (v0.3.1) |
| Jev 1.13.0 | Not published | Closed, hosted | 57.96 |
All benchmark figures are self-reported by Nace AI. The company also reports a 97.3 score on InfographicVQA and 81.9 in the Tools area of its index. Treat those as the maker’s claims until independent testing lands — but the weights are public, so anyone can check.
Why it matters
Decision models are becoming a distinct layer in agent stacks: Cloudflare, AWS and others have shipped small, fast models whose only job is to pick, route or approve. Most of the capable ones have been hosted-only. An open-weight 27B model that runs on a 24 GB Mac changes the privacy and cost math for teams that want those checks on-premise — the hosted version lists a launch price of $0.03 per 1,000 decisions, and the local version costs nothing per call. For builders, this is worth a test download; for everyone else, it is another sign that “AI model” increasingly means a specialist, not just a bigger chatbot.
Frequently asked questions
What is a decision model?
A model that returns probabilities for a defined set of options instead of generating text. You give it context and questions; it gives you a scored choice in one pass.
Can Drex Max v1 run on a Mac?
Yes. The affine Q4 checkpoint targets Apple Silicon Macs with 24 GB or more of memory, and the Q8_0 GGUF targets 48 GB machines. Local Mac paths are text-only for now.
Is Drex Max v1 free to use?
The weights are public under Nace AI’s Open RAIL-M license. Check the license terms for your use case before shipping it in a product.
Sources: MarkTechPost, Wisevoter