Nace AI Open-Sources Drex Max v1, a 27B Decision Model for Macs

Nace AI has released Drex Max v1 as open weights under the Nace.AI Open RAIL-M license, giving developers a 27-billion-parameter decision model that can run entirely on a Mac instead of a data-center GPU cluster. According to the official model documentation, the release ships bf16 weights, a Q8_0 GGUF checkpoint at 27.25GB, an affine Q4 build at 14.45GB, and forks of llama.cpp and Ollama, all serving the same POST /v1/systemone API so routing, classification and policy checks can stay on local hardware.

What Drex Max Is

Drex Max is a dense 27-billion-parameter model with a 262,144-token context window (16,384 by default) that answers typed questions about a given state of text, JSON or images, returning a probability for every option in a single forward pass rather than generating text. The architecture uses 64 layers, with 48 Gated DeltaNet linear-attention layers running left to right and 16 full-attention layers placed every fourth layer for non-causal, whole-sequence attention. A 255-way readout head converts the final hidden state into option probabilities, and the model handles three question types: named-option choice, yes/no, and ordinal scoring.

According to Nace's self-reported benchmarks on the Decision Index 0.4, Drex Max scores 67.52, which is 6.27 points ahead of Nace's own Drex v2.0 (61.25) and ahead of Jev 1.13.0 (57.96). The model also achieves 97.3 skill on InfographicVQA, 73.92 on the public vision index, and 81.9 on the Tools benchmark area. On the Artificial Analysis Decision Index, Nace reports Drex Max ranks fourth of 18 at a score of 64, behind three models tied at 65 and ahead of Claude Haiku 5.5 (Low) at 61.

Deployment Paths and Hardware Requirements

Drex Max runs across five distinct runtime environments, all serving the same API. On NVIDIA hardware, the model runs in full Python on an H100 in bf16 (approximately 49GB loaded) or on an L40S 48GB card in 8-bit (approximately 26GB loaded). For Apple Silicon, Nace provides four native paths: llama.cpp, Ollama, the Kapitsa runtime, and Apple's MLX framework.

The affine Q4 checkpoint, at 14.45GB, targets Macs with 24GB of memory or more and runs via Kapitsa or MLX. The Q8_0 GGUF checkpoint, at 27.25GB, requires 48GB of unified memory on Apple Silicon. Nace recommends a 2048-token context setting for 48GB Macs running its Kapitsa runtime. For Ollama users, setting num_gpu to 40 layers enables Metal GPU offload on a 48GB Mac, while num_gpu 0 forces CPU-only inference. Local Mac and CPU paths are text-only for now; image inputs still require the Python path with the native vision tower.

According to the official model documentation, tested hardware includes an Apple M5 Pro with 48GB unified memory running macOS 26.5.2, where the Q8_0 GGUF on Metal answered requests within 0.02 per probability of the bf16 weights run on an NVIDIA H100.

Quantization Fidelity and Accuracy Claims

The quantization approach is a standout technical claim. According to Nace's self-reported testing documented on the model page, on a 48GB Apple M5 Pro, the Q8_0 GGUF matched H100 bf16 output within 0.02 per probability, and the affine Q4 build stayed within 0.0329 of Q8_0 with no choice flips. These figures suggest that the quantization strategy preserves decision accuracy across different hardware platforms and precision levels.

On the Decision Index 0.4 benchmark suite—which includes 140,178 scoreable requests across 37 benchmarks in five areas (Knowledge & Reasoning, Language, Retrieval, Tools, and Arts)—Drex Max v1 self-reported results show 75.4 on the raw metric and 67.52 on the public index. In the Tools area specifically, the model scores 81.9, and on the vision benchmarks, InfographicVQA yields 97.3 skill. The model achieves 92.2% on JevBench (213/231 correct) and 94.30% on Decision Bench v4 (1010/1071 correct).

Why This Matters: MachineHours Analysis

MachineHours conditional analysis: The release of Drex Max at 27B parameters with demonstrated quantization fidelity challenges one key assumption in the current AI infrastructure debate. A dense decision model that can match H100-class accuracy on a consumer 48GB Mac, and run acceptably on 24GB entry-level Apple Silicon, shifts the cost-benefit calculation for narrow inference workloads. Decision tasks—routing, ticket triage, guardrail checks, moderation flags—represent a large fraction of cloud API call volume. If Nace's self-reported benchmark numbers hold under independent scrutiny, then workloads in this category no longer require cloud GPU capacity for latency or accuracy reasons. That does not replace data-center training runs or eliminate GPU infrastructure for model development, but it does reallocate inference economics at the edge. The 67.52 Decision Index 0.4 score, achieved at 27B parameters on quantized weights, also demonstrates that parameter count alone does not determine decision-task performance; a specialized architecture with non-causal attention and a direct readout head can outperform much larger generative models on choice and classification tasks. For context, Drex v2.0 at 35B scores 61.25 on the same benchmark, suggesting that Drex Max's architecture innovation, not scale, drives its performance gain.

License and Production Readiness

The weights are released under the Nace.AI Open RAIL-M license. Specific usage terms should be checked against the license text. Enterprises evaluating whether they can fine-tune or redistribute derivative decision models built on the 27B backbone should consult the license terms and the GitHub repository for usage examples and API specifications.

Further Reading and Context

Drex Max's open-weight release extends a pattern that MachineHours has tracked since Stable Diffusion's 2022 release broke the monopoly of closed API image models, and since Meta's Llama 4 pushed permissive multimodal weights into frontier territory. MachineHours' own analysis of the open-source-versus-closed-AI battle argues this dynamic reshapes who captures value in AI infrastructure. Readers tracking multimodal reasoning more broadly may also find relevant background in MachineHours' coverage of GPT-4 Multimodal, though Drex Max is a narrower, non-generative decision architecture rather than a general chat model.

What to Watch

The near-term questions are adoption and validation. Does the 24GB Mac minimum spec actually pull routing and classification workloads off cloud infrastructure, or does the text-only limitation on local runtimes (images still require the Python/NVIDIA path) keep most production deployments server-side? Watch whether independent benchmarks beyond Nace's self-reported Decision Index and public-suite results confirm the quantization fidelity claims. The Artificial Analysis Decision Index ranking at 64 points provides one external reference, though broader third-party benchmarking would strengthen confidence. Finally, monitor whether Nace publishes deployment case studies showing latency, throughput, and cost comparisons between local Mac inference, CPU-only paths, and cloud GPU endpoints, since the decision-model economics depend heavily on real-world deployment patterns, not just synthetic benchmarks.

The affine Q4 checkpoint documentation details the developer-reported quantization measurements.