Drex Max is built for decisions. Nace AI has released Drex Max v1 as open weights, presenting it as the company’s most capable decision model and a 27B system that can run on Apple Silicon and other hardware. This article is sponsored by Nace AI.
Unlike a generative model, Drex Max answers typed questions about a given context. A request can include a state in text, JSON, or images, plus named questions; the model then returns a probability for every option in one forward pass. Nothing is generated.
Questions come in three forms: a choice among named options, a yes-or-no question, or an ordinal score. That design targets agent routing, ticket triage, classification, and guardrail checks—tasks where a clean probability is more useful than another paragraph of synthetic confidence.
A 27B Architecture Built Around Questions
The model card describes a dense 27B backbone with 64 layers. Forty-eight are Gated DeltaNet linear-attention layers that run left to right, while the remaining 16 are full-attention layers placed every fourth layer. Those full-attention layers are non-causal, so every token can attend to the whole sequence.
Images pass through a native vision tower and remain visible to every question in a request. A 255-way readout head converts the last hidden state into option probabilities, allowing up to 255 options per question. The configuration lists a qwen3_5 model type and a 262,144-position window, with 16,384 listed as the default.
Nace’s release includes bf16 weights, a Q8_0 GGUF, an affine Q4 checkpoint, and forks of llama.cpp and Ollama. The package also supports NVIDIA GPU hardware, including H100 for bf16 and L40S for 8-bit workloads, alongside Apple Silicon paths.
Local Hardware, Quantization, and Results
Nace documents five runtimes, all serving the same POST /v1/systemone API. The affine Q4 checkpoint targets Macs with 24 GB or more, while the Q8_0 GGUF needs 48 GB; a 4-bit build runs on 24 GB Apple Silicon Macs through Kapitsa or MLX.
On a 48 GB Mac, Nace recommends –context 2048 for Kapitsa. Ollama users set num_gpu 40 to offload 40 layers to Metal, or set num_gpu 0 to run on the CPU. The Ollama fork also needs a CAPABILITY decision line in the Modelfile.
The available files cover several memory targets: bf16 at about 49 GB, 8-bit at about 26 GB, Q8_0 GGUF at 27.25 GB, and affine Q4 at 14.45 GB. Local Mac and CPU paths are text only for now, so the vision capability does not follow every deployment route.
Quantization held up in Nace’s tests. On a 48 GB Apple M5 Pro, the Q8_0 GGUF matched H100 bf16 within 0.02 per probability, while affine Q4 stayed within 0.0329 of Q8_0 with no choice flips. That is the useful kind of compression result: smaller files without changing the selected answer.
All figures below are self-reported by Nace AI. Drex Max scored 67.52 on Decision Index 0.4 for public text, ahead of Drex v2.0 at 61.25 and Jev 1.13.0 at 57.96. The listed area scores were Tools 81.9, Language 72.2, Retrieval 68.8, Knowledge and Reasoning 57.1, and Arts 53.4.
The public vision index reached 73.92 across 11 benchmarks and 12,210 rows, with InfographicVQA listed at 97.3. Artificial Analysis Decision Index scored Drex Max at 64, which Nace says ranks 4th of 18. JevBench reached 92.2% (213/231), while Decision Bench v4 reached 94.30% (1010/1071).
The results dataset covers 140,620 requests, and unanswered requests count as wrong. That detail belongs next to the scores, not in the fine print—benchmark numbers without evaluation rules are just decorative furniture.
Drex Max v1 weights are public under Nace.AI Open RAIL-M. The hosted version is presented for routing, classification, and policy checks, while the local release gives developers a path to keep those workloads on a local machine. For regulated industries such as healthcare, legal, and financial services, on-premises or company-cloud deployment can make AI viable when keeping data and customization in-house matters.
Open-weight is not the same as open source: users typically receive the weights, but not the training data, the code, or a real license. Nace’s release makes the model available under Open RAIL-M, but the wider distinction still matters when a business evaluates control, safety, and long-term access.
Based on
- Nace AI Open-Sources Drex Max v1: A 27B Decision Model That Runs Locally on Apple Silicon — marktechpost.com
- Open-weight AI models are getting cheaper and better. Here’s how to use them without taking on the cost of running them | VentureBeat — venturebeat.com
- ZenMux Introduces Kimi K3.1 Pro for Long-Context AI Workflows – USA Today — usatoday.com
- Vibe coding is old news. People are ‘vibe manufacturing’ now. — africa.businessinsider.com
