The short version
- Open-weight 27B-parameter multimodal decision model released by Nace AI on Hugging Face (`nace-ai/drex-max-v1`).
- Scores and ranks up to 255 options in a single forward pass rather than generating free-form chat text, achieving 67.52 on Decision Index 0.4.
- Accepts text, structured JSON, and image inputs with a 262K token context window.
- Runs 100% locally on Apple Silicon M4 and M5 Macs with 24GB+ unified memory, with support for Ollama, llama.cpp, Apple MLX, and Unsloth.
At a glance
| Developer | Nace AI |
|---|---|
| Parameters & context | 27B parameters, 262K token context window |
| Modalities | Text, structured JSON, and images |
| Decision capacity | Scores up to 255 options in 1 forward pass (Decision Index 0.4: 67.52) |
| Local runtimes | Ollama, llama.cpp, Apple MLX, Unsloth (24GB+ M4/M5 Mac) |
| Category | AI & Models |
| Creator / publisher | Markandey Sharma |
| Official link | huggingface.co ↗ |
| Heat | 🔥 33.9K Heat |
| Announced | |
| Data updated |
Why score 255 options in a single forward pass instead of autoregressive chat
Standard large language models make decisions by generating tokens one by one—often requiring long chain-of-thought prompts and fragile JSON parsing just to pick the best route, tool call, or candidate item. Nace AI built Drex Max v1 (`nace-ai/drex-max-v1`) specifically as a non-generative decision engine.
Given a prompt consisting of text, structured JSON schemas, and images across up to 262,000 tokens of context, the 27B model evaluates up to 255 candidate options in a single forward pass and outputs calibrated scores for each option without autoregressive decoding overhead.
Running a 27B multimodal judge locally on 24GB+ M4 and M5 Macs
Because Drex Max v1 only needs a forward pass to score options rather than streaming hundreds of output tokens, it runs efficiently on consumer workstations. Quantized checkpoints run locally on Apple Silicon M4 and M5 Macs equipped with 24GB or more of unified memory.
Out of the box, the weights integrate with Ollama, llama.cpp, Apple MLX, and Unsloth, making it straightforward to drop a private reranker, agent router, or multimodal classifier into local workflows.
Benchmark performance and open-weight license terms
On Decision Index 0.4, Drex Max v1 posts a score of 67.52 across multi-option structured evaluation tasks. While the weights are free to download from Hugging Face for research and standard development, Nace AI's license includes restrictions against high-risk automated decision-making in regulated domains.
How to try it
Open nace-ai/drex-max-v1 on Hugging Face
Use the button above to inspect the model card, benchmark tables, and quantized weight files.
Load the model in Ollama, MLX, or llama.cpp
Download the quantized checkpoint suited for your GPU or 24GB+ M4/M5 Mac.
Pass text, JSON, or images with up to 255 options
Run a single forward pass to get ranked option scores for routing, classification, or reranking.
FAQ
What makes Drex Max v1 different from a regular 27B LLM?
Instead of generating conversational text token by token, Drex Max v1 is a multimodal decision model that scores up to 255 candidate options in a single forward pass.
Can I run Drex Max v1 locally on a Mac?
Yes. Quantized versions run locally on Apple Silicon M4 and M5 Macs with 24GB or more of unified memory via MLX, Ollama, or llama.cpp.
How large is the context window?
Drex Max v1 supports a 262K token context window across text, JSON, and image inputs.
