Nace.AI said on X on 30 October that it is “open-sourcing Drex-Max v1” and that the model “runs 100% locally on Apple Silicon (M4 & M5)”. The weights are now public on Hugging Face under the Nace.AI Open RAIL-M license, which does not cover commercial use by any company with more than $1 million in revenue over the prior twelve months or more than $1 million in funding raised.
By our arithmetic, the 27B-parameter Drex Max scores 6.27 points higher than Nace’s 35B Drex v2.0 on the public Decision Index 0.4 reported on Nace’s model card, with 8 billion fewer parameters, about 23% fewer.
Business Engineer Course · The FDE Playbook
An evergreen Business Engineer course on forward-deployed engineering, how AI gets put to work inside organisations. It teaches the general idea only and does not discuss Nace or Drex Max.
The key insight: Nace has published the Drex Max v1 weights under an Open RAIL-M license that is free for individuals, research and organizations below $1 million in revenue or funding; by our arithmetic the 27B model scores 6.27 points above the 35B Drex v2.0 on the public Decision Index 0.4 reported on Nace’s card.
Drex Max was already available through Nace’s API, which we covered on 20 October. The new step is the weights. The model card describes Drex Max as “a dense 27B-parameter model with a 255-way readout head”. It answers typed questions about a given context and returns “a probability for every option back in one forward pass”, and the card adds: “Nothing is generated”.
The card gives a 262,144-token context window, with a default limit of 16,384 tokens per request. It says the bf16 weights take about 49 GB.
Decision Index 0.4 skill by area for Nace’s Drex v2.0 (35B) and Drex Max (27B), chance-corrected skill x100, as reported on the Drex Max model card on Hugging Face (scores as of 19 October 2026). The public-index gain of 6.27 points is our arithmetic.
What Runs on a Mac
The card says the model “runs with Python on an NVIDIA GPU”. On Apple silicon, it says, “its text backbone runs on llama.cpp, Ollama, Kapitsa and MLX”. A 4-bit version “runs on Apple silicon Macs with 24 GB of memory or more”, while the 8-bit GGUF file needs 48 GB.
The Mac routes in the card are text-only. For llama.cpp it says “images need the Python path”, for the GGUF file “Images are not in the GGUF”, and for MLX “Text only”. Nace’s X post describes Drex Max as a multimodal model that runs locally on Apple silicon; the card’s Mac instructions cover its text side.
The License Line
The license defines “Open” as freely accessible “including for commercial use by individuals and smaller organizations below the thresholds in Attachment A”. Attachment A excludes any user whose organization “generated more than one million US dollars ($1,000,000) in gross revenue in the prior twelve (12) months” or “has raised more than one million US dollars ($1,000,000) in total equity or debt funding from any source”, except for personal use or research.
It also excludes anyone whose organization offers a product that “competes with any product or service offered or made available by Licensor or its affiliates, including decision models and document intelligence services”. Use above those thresholds “requires a separate written commercial license from Licensor”, the license says.
How It Scores
Nace’s card reports 67.52 on the public Decision Index 0.4, against 61.25 for Drex v2.0. Drex Max leads in four of the five areas the card lists: tools, language, retrieval, and knowledge and reasoning. It is 0.5 points lower on arts, at 53.4 against 53.9.
On the Artificial Analysis Decision Index, the card says Drex Max scores 64 and “ranks fourth of 18”, one point behind three models at 65. That is 3 points above Claude Haiku 5.5 (Low) at 61 and 9 points above GPT-6 Luna (Low) at 55, by our arithmetic from the card’s table.
Why It Matters
Drex Max is a decision model, not a chatbot: the card says it returns a probability for each option rather than generating text, as in its example of routing a support ticket. Our recent coverage has also followed Samaya’s Sterling model and Figma’s use of Llama. Nace earlier said its Drex search cut coding agents’ steps by 62%.
The Structural Read
As we read the license, the open part is the weights and the free tier; commercial use by organizations above the $1 million revenue or funding thresholds, and by competitors, needs a separate written license from Nace.
As we read the model card, the local Apple silicon routes it documents are for the text backbone, from 24 GB of memory with the 4-bit version, while image inputs use the Python path on an NVIDIA GPU.
As we read the benchmark table, the gain came with 8 billion fewer parameters than Drex v2.0, about 23% fewer by our arithmetic, in four of the five areas the card lists.
Nace.AI, Drex Max v1 license (Open RAIL-M)
“including for commercial use by individuals and smaller organizations below the thresholds in Attachment A”
Three Implications
FOR SMALL TEAMS The license grants free commercial use to individuals and organizations below $1 million in revenue and below $1 million in funding, the license says.
FOR LARGER COMPANIES Organizations above either threshold need a separate written commercial license from Nace, according to the license.
FOR MAC DEPLOYMENTS The model card says the 4-bit version runs on Apple silicon Macs with 24 GB of memory or more, for text inputs.
The framework says “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
The framework is the owner’s general argument about technology stacks, not evidence about Nace; it does not discuss Drex Max or its license.
What Is Not Established
All benchmark figures come from Nace’s own model card, with scores as of 19 October 2026. The Decision Index and Artificial Analysis figures are as the card reports them. The card’s tested hardware lists NVIDIA H100 and L40S runs and, for its text-only Mac route, an Apple M5 Pro with 48 GB of memory, where it says the 8-bit file answers its example request “within 0.02 of the bf16 weights run”. The card gives no speed figures for Apple silicon and does not mention an M4 chip.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Nace has put Drex Max’s weights on Hugging Face and says the model runs locally on Apple silicon, where its card documents text-only routes from 24 GB of memory. The license is free for individuals, research and small organizations, while companies above $1 million in revenue or funding, and competitors, need a separate deal with Nace.
A note on sourcing. Model, benchmark and hardware details come from Nace.AI’s Drex Max model card on Hugging Face and its license file, read on 30 October 2026; benchmark scores are Nace’s own. Figures labelled our arithmetic are our own calculations.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
