Nace AI open-sources Drex Max v1: its 27B decision model now runs on your own hardware

Nace AI has released the open weights of Drex Max v1, its most capable decision model, and the flagship of its Drex family can now run on a developer's own hardware. Until now, Drex Max was available only through Nace's hosted API: when the model launched on October 19, no open weights or quantized builds had been announced. The weights release, reported on November 1, changes where the model's decisions can happen — from Nace's console to a local GPU or Mac.

The release ships in several forms: bf16 weights, a Q8_0 GGUF, an affine Q4 checkpoint, and forks of llama.cpp and Ollama, according to the Hugging Face model card. The weights are released under the Nace.AI Open RAIL-M license. All five documented runtimes serve the same request format as the hosted API, so code written against the hosted model keeps working locally.

A model that scores options instead of writing text#

Drex Max does not generate text. It takes a context — text, JSON or images — plus named, typed questions, and returns a probability for every option in a single forward pass. Questions come in three types: a choice among named options, a yes/no answer, and an ordinal score. That makes it a fit for the small, repeated decisions inside software: agent routing, ticket triage, tool selection and policy or guardrail checks.

The architecture behind those scores is a dense 27-billion-parameter backbone with 64 layers: 48 Gated DeltaNet linear-attention layers that run left to right, and 16 full-attention layers, placed every fourth layer, that are non-causal, so every token can attend to the whole sequence. Images pass through a native vision tower and are visible to every question in a request. The model takes up to 255 options per question, with a default limit of 16,384 tokens per request that can be raised to the full 262,144-token window.

The scores Nace reports — and their date#

Every benchmark figure in this story is self-reported by Nace AI, and the model card dates the performance section as of October 19, 2026 — the scores predate the weights release and describe the same v1 model. On the public Decision Index 0.4, Drex Max scores 67.52, against 61.25 for Drex v2.0 and 57.96 for Jev 1.13.0, over a run of 140,620 requests in which unanswered requests counted as wrong. Its public vision index is 73.92 across 11 benchmarks and 12,210 rows, including 97.3 on InfographicVQA, and its Tools area skill reaches 81.9. Nace also reports 64 on the Artificial Analysis Decision Index, fourth of 18 systems, 92.2% on JevBench (213 of 231) and 94.30% on Decision Bench v4 (1,010 of 1,071).

One confusion is worth heading off. Nace's Drex product page now promotes a newer Drex Max v1.1 with higher reported scores. The open weights are v1, and the figures above are v1's — do not mix the two versions' numbers.

Five ways to run it — with a text-only catch on Macs#

Nace documents five runtimes. Python on an NVIDIA GPU runs the full model, including images: the bf16 weights take about 49 GB, and loading the backbone in 8-bit takes about 26 GB, tested on an H100 and a 48 GB L40S. The llama.cpp and Ollama forks run the text backbone on Apple Metal and CPUs using the Q8_0 GGUF, a 27.25 GB file that needs a 48 GB Mac. Kapitsa version 0.2.25 or later runs the Q8_0 GGUF or the affine Q4 checkpoint on Apple Silicon, and MLX runs the affine Q4 checkpoint — a 14.45 GB model folder for Macs with 24 GB of memory or more.

Quantization holds up in Nace's tests. On a 48 GB Apple M5 Pro, the Q8_0 GGUF matched the H100 bf16 run within 0.02 per probability, and the affine Q4 checkpoint stayed within 0.0329 of the Q8_0 run, with no choice flips. The catch is images: the local Mac and CPU paths run the text backbone only, so screenshots and photos still need the Python GPU path. The hosted API remains the route for image-heavy computer-use work at scale, while the local builds suit text decisions that should stay on the user's own machine.

For teams already using Drex 2.0 or 1.5 locally, the practical change is scope: the most accurate model in the family, and the only one that reads images, is no longer confined to Nace's API. Routing, classification and policy checks that ran as API calls can now run offline — with the same typed answers, and with the same caveat that applies to every vendor-reported benchmark in this story: nobody outside Nace has independently reproduced these scores on the newly released weights yet.

Sources: Nace AI, Drex Max v1 model card on Hugging Face (primary source); MarkTechPost, November 1, 2026; Wisevoter, November 1, 2026; Nace AI, Drex product page; MarkTechPost, October 19, 2026 (hosted launch) — checked November 1, 2026. Benchmark figures are self-reported by Nace AI.