Nace AI released Drex Max v1 as open weights, optimized for local runs including Apple Silicon. The decision model packs 27B parameters and supports a 262,144-token window (16,384 by default). For each query, it returns probabilities for up to 255 options with no text generation. Routing, classification, and policy checks can now stay on your own hardware, offline.
What is Drex Max v1?
Drex Max v1 is a decision model from Nace AI that answers typed questions about a given context. You pass a state (text, JSON, or images) plus named questions. The model returns a probability for every option in one forward pass—nothing is generated.
It supports three question types: choice among named options, yes/no, and an ordinal score. This makes it a fit for agent routing, ticket triage, and guardrail checks. The hosted version is presented at routing, classification, and policy checks.
Drex Max v1 outputs a fixed-size probability vector—up to 255 options per question. That aligns well with pipelines that require crisp, non-generative decisions.
How does Drex Max v1 work?
The model card describes a dense 27B backbone with 64 layers. 48 are Gated DeltaNet linear-attention layers running left to right. The other 16 are full-attention, placed every fourth layer and non-causal: every token can attend to the whole sequence.
Images go through a native vision tower and are visible to every question in a request. A 255-way readout head turns the last hidden state into option probabilities. The config lists a qwen3_5 model type with a 262,144-position window.
This blend of linear and full attention yields a long window, while handling text, JSON, and images in a single request.
How to run Drex Max v1 locally?
The release ships bf16 weights, Q8_0 GGUF, an affine Q4 checkpoint, and forks of llama.cpp and Ollama. You can run it on NVIDIA GPUs, Apple Silicon, and even CPU via llama.cpp and Ollama.
On NVIDIA, use Python: target an H100 in bf16 or an L40S 48 GB in 8-bit. On Apple Silicon Macs, use 24 GB+ for affine Q4 or 48 GB for Q8_0 GGUF. CPU inference is available via llama.cpp and Ollama thanks to the provided weight formats.
Practically: choose a weight artifact (bf16, Q8_0 GGUF, or affine Q4) for your hardware, set up Python or llama.cpp/Ollama, pass in state (text/JSON/images) and named questions—get option probabilities in one pass.
Drex Max v1 performance and benchmarks
Drex Max v1 scores 67.52 on the public Decision Index 0.4 and is 6.27 points ahead of Drex v2.0 (self-reported). It posts 97.3 skill on InfographicVQA and reaches 81.9 in the Tools area. This is a top-tier decision model you can now run offline on a Mac.
The numbers span overall decision quality plus specialized tracks for visual reasoning and tools. The long window and 255-way readout are a strong match for routing and guardrails.
Why local Drex Max v1 matters for businesses
Open weights mean routing, classification, and policy checks can stay on a local machine. You can run it offline on a Mac and keep execution on your own hardware.
The 27B architecture with a 262,144-token window and a non-generative probability head plugs neatly into verification and ticket-triage pipelines. With support for Apple Silicon, NVIDIA GPUs, and CPU via llama.cpp/Ollama, it’s a versatile on-device choice.
| Parameter | Drex Max v1 | Drex v2.0 | OpenAI models | Google's decision models | Anthropic's AI models |
|---|---|---|---|---|---|
| Parameters | 27B | — | — | — | — |
| Context window (max) | 262,144 | — | — | — | — |
| Default context window | 16,384 | — | — | — | — |
| Options per question | up to 255 | — | — | — | — |
| Decision Index 0.4 | 67.52 | — (v1 +6.27) | — | — | — |
| InfographicVQA | 97.3 | — | — | — | — |
| Tools area | 81.9 | — | — | — | — |
| Local on Apple Silicon | available | — | — | — | — |
FAQ
What is Drex Max v1 and how does it work?
It’s a decision model from Nace AI that takes a state (text, JSON, or images) and answers named questions by returning option probabilities in a single forward pass. It supports three question types: named-option choice, yes/no, and an ordinal score. The model has 27B parameters and a 262,144-token window (16,384 by default), operating without text generation.
Drex Max v1 vs Drex v2.0 comparison — what’s the gap?
On the public Decision Index 0.4, Drex Max v1 scores 67.52 and is 6.27 points ahead of Drex v2.0 (self-reported). No other differences are provided in the source. If this metric is critical, validate it for your target tasks, as the lead is described as self-reported.
How much does it cost (Drex Max v1 pricing)?
The source states Nace AI released open weights for Drex Max v1 and provided bf16, Q8_0 GGUF, and affine Q4 formats, plus forks of llama.cpp and Ollama. No pricing is provided. You can run locally on NVIDIA GPUs, Apple Silicon, or CPU via llama.cpp/Ollama using the available weights.
Drex Max v1 for Apple Silicon teams — who is it for?
Teams on Apple Silicon Macs. For affine Q4 you need 24 GB+ memory, for Q8_0 GGUF you need 48 GB. The model runs offline and returns probabilities without generation, fitting routing, classification, and policy checks on-device. NVIDIA GPU and CPU paths are also available.
How to get started locally with Drex Max v1?
Select a weight format: bf16 (for H100 in bf16), Q8_0 GGUF (for 48 GB on Apple Silicon), or affine Q4 (for 24 GB+ on Apple Silicon). Then use Python on NVIDIA, or llama.cpp/Ollama for CPU/Mac. Pass state (text, JSON, or images) and named questions—receive probabilities over up to 255 options.
Based on MarkTechPost.
