{"slug": "nace-ai-releases-drex-technical-report-and-27b-drex-max-decision-model", "title": "Nace AI releases Drex technical report and 27B Drex Max decision model", "summary": "Nace AI published an arXiv technical report titled \"Towards Foundation Decision Models: Drex\" around November 5, 2026, alongside Drex Max, a 27-billion-parameter decision model that returns calibrated choices in a single forward pass rather than generating text. Nace AI says Drex Max reaches a median response time of approximately 183 milliseconds, versus 2 to 3.5 seconds for comparable systems, and scored 67.52 on Decision Index 0.4, which the Palo Alto startup calls a new state of the art among decision models. The full Drex family, including a 9B variant, ships under the Nace.AI Open RAIL-M license via GitHub and Hugging Face; Nace AI raised a $21.5 million seed round led by Walden Catalyst in May 2026.", "body_md": "# Nace AI releases Drex technical report and 27B Drex Max decision model\n\nThe Palo Alto startup's flagship model scores 67.52 on Decision Index 0.4 with a median latency of roughly 183 milliseconds\n\nMost AI models answer a question by writing. Nace AI wants its models to simply decide.\n\nThe Palo Alto startup has published a technical report titled “Towards Foundation Decision Models: Drex” on arXiv, released around November 5, 2026. Alongside it comes Drex Max, a 27-billion-parameter model built to pick from a set of options rather than generate paragraphs of text.\n\nThe pitch is speed and cost. Nace AI says Drex Max hits a median response time of approximately 183 milliseconds. Comparable systems, by the company’s measure, take between 2 and 3.5 seconds to respond.\n\n## What Drex Max actually does\n\nTechnically, the model returns calibrated outputs tied to those user-defined choices. It does this in a single forward pass instead of using the autoregressive approach common to chat-style models.\n\nThe model is also multimodal. Drex Max accepts text, images, JSON and video as inputs. It supports a 262,000-token context window.\n\nOn benchmarks, Drex Max scored 67.52 on Decision Index 0.4. Nace AI describes that result as a new state-of-the-art among decision models.\n\n## How the models were trained\n\nThe technical report centers on how Nace AI trained the Drex family. The company used a method it calls RLAF, combined with supervised fine-tuning and consistency losses.\n\nDrex Max is not the only release. The full Drex model family is available, including a smaller 9B variant. The models and associated resources are being distributed through GitHub and Hugging Face.\n\n### AI, tech, and the markets they move—in one daily briefing.\n\nDaily. Free. Join 34,000+ readers across crypto, finance, and policy.\n\nThe family ships under the Nace.AI Open RAIL-M license.\n\n## Background on Nace AI\n\nNace AI is a young company. It was founded in 2024 and is based in Palo Alto.\n\nIn May 2026, the startup raised $21.5 million in seed funding. Walden Catalyst led the round, with General Catalyst also participating.\n\nThe company has already been packaging its decision technology into specific products. Those include Drex for Search and Drex for Finance.\n\n## What this means\n\nThe interesting bet here is architectural. Much of the AI industry has spent the past few years making models bigger and more conversational. Nace AI is arguing that a large share of enterprise work does not need conversation at all.\n\nCalibration is the quieter selling point. In enterprise settings, knowing how confident a model is can matter as much as the answer itself.\n\nThe product lineup hints at the target customers. Drex for Finance suggests banks and financial firms, where latency and auditability both carry weight. Drex for Search suggests ranking and retrieval.\n\nThe open-release strategy is the other lever. Putting the full family, including the 9B variant, on GitHub and Hugging Face lowers the barrier for developers to test the approach.\n\nThe latency and benchmark comparisons come from Nace AI itself. Decision Index 0.4 is also a specialized benchmark. The license terms will shape adoption too, as Open RAIL-M style restrictions can limit some commercial uses.\n\n**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).", "url": "https://wpnews.pro/news/nace-ai-releases-drex-technical-report-and-27b-drex-max-decision-model", "canonical_source": "https://cryptobriefing.com/nace-ai-drex-max-decision-model/", "published_at": "2026-11-06 14:34:10+00:00", "updated_at": "2026-11-06 14:50:24.532232+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "ai-research", "ai-startups", "large-language-models"], "entities": ["Nace AI", "Drex Max", "Drex", "Decision Index 0.4", "Walden Catalyst", "General Catalyst", "Hugging Face", "GitHub"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/nace-ai-releases-drex-technical-report-and-27b-drex-max-decision-model", "markdown": "https://wpnews.pro/news/nace-ai-releases-drex-technical-report-and-27b-drex-max-decision-model.md", "text": "https://wpnews.pro/news/nace-ai-releases-drex-technical-report-and-27b-drex-max-decision-model.txt", "jsonld": "https://wpnews.pro/news/nace-ai-releases-drex-technical-report-and-27b-drex-max-decision-model.jsonld"}}