Overview
Drex Max is a decision model from Nace AI. It only decides, never writes. You hand it a state, as text, JSON, a screenshot or video, plus a set of named questions, and it returns a probability for every option in a single forward pass without generating a single token.
The line now splits in two. The version with public weights is 27B dense: 64 layers, of which 48 are Gated DeltaNet linear attention and 16 are full attention placed every fourth layer with no causal mask, so every token sees the whole sequence. The window is 262,144 positions with 16,384 as the default, and a single question can carry up to 255 options. The hosted v1.1 is what the official site leads with, scoring 70.8 on the public Decision Index with a 172 ms median, aimed at driving computer use from screenshots.
The useful place for a model like this is not chat but the forks inside an agent: which element to click, which tool to call, who owns this ticket, whether a reply clears a guardrail. Those decisions used to be pushed into a large model as a side effect of generation, or handed to a small classifier trained in house. Drex Max takes the middle path, with a probability distribution as output and a fixed interface, so latency and cost can be estimated up front. With the weights public, a 4-bit build runs on a Mac with 24GB of memory.
Key Features
- Probabilities instead of text: Every named option gets a probability directly. Choice, yes or no, and ordinal scoring questions share one interface, and several questions can ride in a single request.
- Long context: A 262,144 position window with 16,384 as the default, enough to place a whole document or a screen of structured state before asking.
- Images and screens: A native vision tower makes images visible to every question in the request, covering screenshot clicks, chart reading and infographic questions. That path runs on the Python runtime with an NVIDIA GPU.
- Local deployment: Public weights come as bf16, Q8_0 GGUF and affine Q4 checkpoints. The 4-bit build is about 14.45GB and runs on an Apple silicon Mac with 24GB or more, while Q8_0 needs 48GB.
- Five runtimes: Python, llama.cpp, Ollama, Kapitsa and MLX all serve the same POST /v1/systemone endpoint, so moving hardware does not change calling code.
- Open weights and hosted side by side: The 27B model is released under Nace.AI Open RAIL-M for self-hosting, while the stronger generation runs on OpenRouter and the official console behind the same calling convention.
Use Cases
- Tool routing inside an agent: treat the candidate tools as options and ask which one this turn should call.
- Computer use: read one screenshot and decide which element to click next.
- Ticket triage: read title, body and history, then decide the owning team, priority and whether a person is needed.
- Guardrail checks: score whether a drafted reply crosses a policy line before it is sent.
- Document and contract decisions: read a clause or a report and decide whether a conclusion holds.
- Air-gapped work: keep model and data on an internal machine so routing and classification never leave the network.
Pros
- One forward pass per request, so latency and unit cost can be estimated before deployment.
- Probabilities rather than prose make thresholds and fallback paths straightforward to code.
- It can only answer inside the options you supply, so a fourth answer cannot be invented.
- Quantisation holds up: Q8_0 on an M5 Pro stayed within 0.02 per probability of H100 bf16.
- A 262K window takes long documents whole instead of chunking first and stitching answers after.
- Open weights let routing and compliance decisions stay entirely on local hardware.
Pricing
Public weights are free to download under Nace.AI Open RAIL-M. Hosted usage is metered per decision: Drex Max v1.1 costs 0.03 dollars per thousand decisions at its launch price, which is half the regular rate, against 0.06 for Drex 2.0 and 0.04 for Drex 1.5. Self-hosting depends on hardware: bf16 needs about 49GB of GPU memory on an H100, 8-bit fits an L40S at 48GB, and on Apple silicon Q8_0 needs 48GB of memory while affine Q4 needs 24GB or more.
Summary
Drex Max narrows the job to one thing: given a state and questions, return option probabilities. Narrow is the point, because cost and latency become calculable, which suits a model called at every fork in an agent rather than once in a conversation. Its most concrete value sits in computer use, where one screenshot, one decision and a little over a hundred milliseconds beat a general model thinking out loud by an order of magnitude. The public 27B weights score 67.52 on Decision Index 0.4, while the hosted v1.1 the site now leads with scores 70.8 on the same suite. Before adopting it, check whether your question can be written as a finite set of options. If it can, this is steadier than letting a large model improvise; if it cannot, the model has nothing to offer.
Version History
- Drex Max v1 open-sourced: a 27B decision model that runs locally (2026-11-01): Nace AI released Drex Max v1, its flagship decision model, under Nace.AI Open RAIL-M. It is 27B dense with 64 layers, a 262,144 token window and up to 255 options per question, reading text, JSON or images and returning option probabilities in one forward pass. Public Decision Index 0.4 scores 67.52, 6.27 points above Drex 2.0 at 61.25. bf16, Q8_0 GGUF and affine Q4 checkpoints ship with llama.cpp, Ollama, Kapitsa and MLX builds; the 4-bit checkpoint is about 14.45GB and runs on a 24GB Apple silicon Mac. The hosted version on OpenRouter costs 0.03 dollars per thousand decisions, and the official site lists v1.1 at 70.8 on the same index with a 172 ms median.