Nace AI Open-Sources Drex Max, a 27B Model That Answers Without Writing

A different kind of model release

Most model releases this year have been about generating more text, faster, for less money. Nace AI went the other way. On November 1, the company released the open weights of Drex Max v1, a dense 27-billion-parameter decision model that does not write out an answer at all. It returns a probability for every option in a single forward pass, then stops. If you have ever wanted a model that picks instead of rambles, this is built for exactly that job.

The release matters for a practical reason: you can run it yourself. The weights ship in several formats, including bf16 weights, a Q8_0 GGUF file, and an affine Q4 checkpoint, along with forks of llama.cpp and Ollama. Nace AI says the model runs on consumer Apple Silicon Macs and NVIDIA GPUs, which puts it within reach of individual developers and small teams, not just companies with a cluster budget.

What an open weights decision model actually does

A normal chat model generates tokens one by one until it decides it is done. That works well for writing, but it is a clumsy way to make a choice. Sampling settings, stop conditions, and parsing all get in the way, and the same prompt can produce different answers on different runs.

Drex Max skips that whole process. Give it a question with options, and it scores every option in one pass, returning probabilities you can read, compare, and log. The context window is 262,144 tokens, so the options and the evidence behind them can be long: documents, tool results, or retrieved passages. For classification, routing, ranking, and evaluation work, that is a cleaner primitive than asking a chatbot to answer with a single letter and hoping it obeys.

The benchmark numbers give the claim some shape. Drex Max scores 67.52 on the public Decision Index 0.4, which Nace AI says is 6.27 points ahead of Drex v2.0. It also reaches 97.3 on InfographicVQA and 81.9 in its Tools area. Those are vendor-reported figures, so treat them as a starting point rather than a verdict, but the gap over the previous Drex version is large enough to be interesting.

Why open weights change the calculation

Releasing weights under the Open RAIL-M license changes who can use this kind of model and how. A hosted decision API charges per call and keeps your inputs on someone else's servers. A local model costs nothing per call after download, keeps data on your machine, and behaves the same way every time you run it, which matters when you are grading other models or auditing decisions.

That trade has limits. Twenty-seven billion parameters still needs real hardware, and the quantized versions trade some accuracy for size. The Q4 checkpoint will run in more places, but it will not match the full bf16 weights on hard cases. If you are deciding between this and a hosted frontier model, the honest test is to run your own evaluation set through both and compare, not to trust a benchmark table.

The open release also fits a pattern visible across the last few weeks. Mistral opened the weights on its trillion-parameter Large 4, and Google's Gemini 4 Carbon pitched itself to security teams first. Vendors are splitting the market: giant hosted models for general work, and specialized or open models for jobs where control, cost, or privacy matter more than raw scale. Drex Max sits firmly in that second group.

What to try first

If you want to kick the tires, start with a job where you already have ground truth. Take a few hundred examples from your own classification or routing task, run them through Drex Max locally, and measure accuracy against your current approach. That tells you more in an afternoon than any benchmark table will.

Second, test the probability outputs, not just the top pick. A decision model is only useful if its confidence means something. Check whether high-confidence answers are actually right more often than low-confidence ones. If the calibration is good, you can route uncertain cases to a human or a bigger model and automate the rest. If it is not, you have learned something important before building on top of it.

Third, keep an eye on cost per decision. A local 27B model is cheap per call but not free: it needs GPU time and memory. For low-volume, high-stakes decisions, that is a good deal. For millions of trivial classifications a day, a smaller specialized classifier may still win. The right answer depends on your volume, and only your own numbers will settle it.

The bottom line

Drex Max is not trying to be the next general chatbot, and that is the point. It is a focused tool for a common job: look at the options, weigh them, and answer with numbers instead of prose. Open weights, consumer hardware support, and a genuinely different output format make it worth a weekend experiment for any team that routes, ranks, or grades things with models today. The cheaper GPT-6.1 Terra pricing may grab more headlines, but for builders who want control over their decisions, a model you can run yourself is the more useful news.