What happened
Nace.AI released Drex Max v1 on Hugging Face on October 30. The release includes open weights for a dense 27B-parameter model, plus Q8_0 GGUF and affine 4-bit variants. The model is published under Nace.AI's Open RAIL-M licence. The model card says that commercial use is excluded for organizations with more than $1 million in revenue or funding, so availability of weights does not by itself establish broad commercial usability.
Drex Max is presented as a decision model rather than a conventional text generator. A caller supplies a state and a typed question, and the model returns a probability for each available answer. The supported question forms include choice, yes/no, and score. Its 255-way readout head can score up to 255 options in one forward pass. The supplied material describes no free-text generation in that path. Inputs can include text, JSON, and up to eight images, although the local Apple Silicon path described for the release is for text.
The model card lists a maximum context length of 262,144 tokens, with 16,384 tokens as the default. Its architecture has 64 layers: 48 Gated DeltaNet linear-attention layers and 16 non-causal full-attention layers. Nace reports a score of 67.52 on its public Decision Index 0.4, compared with 61.25 for its own Drex v2.0 model. That is a vendor-reported comparison on a vendor index, not an independent evaluation of performance on a particular deployment task.
For local inference, Nace says Drex Max can run on Apple Silicon through llama.cpp, Ollama, Kapitsa, and MLX forks. The affine 4-bit checkpoint requires a Mac with at least 24 GB of memory; the Q8_0 GGUF version requires 48 GB. NVIDIA GPUs running Python are also supported. Image input is described as requiring the Python/NVIDIA route. Nace also offers cloud inference through its console; the supplied release material lists the drex-max API at $0.03 per one million input tokens and says output is not billed.
The immediate application is not replacing a general-purpose assistant. A system can use the model where the action set is known: choose among tools, rank routes, classify a support ticket, or assess a defined checklist. Instead of asking a generative model to explain its choice and then parsing that explanation, a developer can submit the candidate options and consume a bounded probability vector.
What the archive adds
This is a new release in a known decision-model thread. Open d1: Edge decision models for text, vision, and audio covered Liquid AI's open-weight multimodal decision checkpoints. Articul8 launches Bivia, a multimodal decision model offered via API for enterprise agents described an API-oriented enterprise version of the same broader category. Drex Max adds a 27B open-weight release and specified Apple Silicon deployment options.
Trend signal
The release supports Model answers become task interfaces. Drex Max's stated interface is a bounded set of probabilities for typed questions, rather than unstructured language intended for a later parser. The evidence is architectural and product-level, not proof that this interface will outperform generative models in production.
Connections to other events
The archive's Kapitsa v0.2.15, an open-source Apple Silicon inference engine for Qwen and Nemotron, ships stability fixes and faster MoE prefill documented Apple Silicon inference tooling. Drex Max names Kapitsa among its local execution paths, connecting the decision-model release to that local-inference ecosystem.
Why it matters
Agent systems contain many decisions that are smaller than an answer: whether to call search, which specialist to dispatch, which queue should receive a ticket, or whether an extracted record satisfies a fixed rule. A model that returns a score for the available actions can fit those points without a free-form response layer. The 255-option head and one-pass design make that proposition concrete enough to test.
The trade-off is important. Bounded output only helps when a team can define the options and obtain examples for evaluation. It does not solve ambiguous task design, stale policies, or incorrect tool results. The licence also means teams must establish eligibility before treating the released weights as a production component.
Limits and uncertainty
Nace's Decision Index result is self-reported, and the supplied material provides no independent benchmark, error analysis, calibration results, or comparison with a conventional model configured for the same routing task. It also does not establish latency, throughput, or accuracy on a team's labels. The release describes local text inference on Apple Silicon, but says image input requires Python and NVIDIA, limiting the meaning of “local multimodal” for Mac-only deployments. Finally, the licence restriction may rule out commercial use for many prospective adopters.
Skill to build now
Choose one existing agent decision with a fixed action set, such as selecting among five tools or assigning incoming tickets to queues. Create a held-out set of real cases, define the allowed options explicitly, and measure top-choice accuracy plus calibration of the returned probabilities. Run the same cases through the current generative router before changing production flow. This separates the appeal of a bounded interface from evidence that it improves the actual decision.
Reader takeaway
Drex Max v1 is an open-weight, 27B model for choosing among defined options rather than writing an answer. Its Apple Silicon variants make local evaluation feasible on sufficiently large Macs, but the reported benchmark is vendor-run and the revenue-and-funding licence cap must be checked before any commercial deployment.