{"slug": "nace-ai-s-drex-v2-0-bets-agents-don-t-need-words-to-make-a-decision", "title": "Nace.AI's Drex v2.0 bets agents don't need words to make a decision", "summary": "Nace.AI released Drex v2.0 on October 12, a decision model that returns calibrated probabilities for each option in a single pass with no generated text, built on the Qwen3.6-35B-A3B mixture-of-experts architecture with roughly 3 billion active parameters per token and a 131,000-token context window. On the public Decision Index 0.3.1 benchmark, Drex v2.0 scored 61.25, 3.2 points above Nace.AI's Drex v1.5 (released October 9) and 3.3 points ahead of TypeSafe AI's Jev 1.13.0, and it beat Jev 60.9% to its rival in an eight-game arena test while edging it on email classification with an AUC of 0.984 versus 0.981. The Palo Alto startup, which has raised $26.5 million total including a $21.5 million seed round led by Walden Catalyst in May 2026, ships Drex v2.0 under an Open RAIL-M license for local runs via Python, llama.cpp or Ollama.", "body_md": "*Nace.AI just shipped a model that never writes a sentence. It only scores options, and it wants to be the thing every AI agent calls before it acts.*\n\nOn October 12, the Palo Alto startup released Drex v2.0. It's the latest version of a model built for one narrow job: given a situation and a list of choices, return a probability for each option in a single pass, with no prose at all. No chain of thought, no explanation, no tokens spent deciding how to phrase the answer. Just numbers.\n\nThat sounds small until you see where it's aimed. Drex is built to sit inside AI agents at the exact moment they have to pick something: which tool to call, which document is actually relevant, whether a transaction looks fraudulent, whether an email is spam. Those are the decisions that happen thousands of times a second inside any serious agent deployment. Right now most of them get routed through a general-purpose language model that was built to write essays, not flip coins.\n\nDrex v2.0 runs on Qwen3.6-35B-A3B. It's a mixture-of-experts architecture with 35 billion total parameters but only about 3 billion active per token, paired with what Nace.AI calls a Kev pointer head that turns the model's internal state directly into calibrated probabilities. It handles context windows up to 131,000 tokens. It ships under Nace.AI's own Open RAIL-M license, so developers can run it locally through Python, llama.cpp, or Ollama instead of paying per call to a hosted API.\n\nOn the public Decision Index 0.3.1 benchmark, Drex v2.0 scored 61.25. That puts it 3.2 points ahead of its own predecessor, Drex v1.5, which Nace.AI had released just three days earlier on October 9. The bigger number is the 3.3-point lead over Jev 1.13.0, a rival decision model from TypeSafe AI that came out in September and has already built a following among developers building agent routers.\n\n[ElevenLabs launches Eleven v4 Turbo to chase OpenAI in voice agents](https://startupfortune.com/elevenlabs-launches-eleven-v4-turbo-to-chase-openai-in-voice-agents/)\n\nElevenLabs is pricing the trade-off at $0.05 per 1,000 characters on Turbo, half the $0.10 rate for its higher-quality Multilingual v2 and Eleven v3 models. The company is betting enterprises will pay for emotional range, even at double the latency of its Flash model. - [elevenLabs voice agents competing with OpenAI](https://startupfortune.com/elevenlabs-launches-eleven-v4-turbo-to-chase-openai-in-voice-agents/) - [AI phone call latency under 150 milliseconds](https://startupfortune.com/elevenlabs-launches-eleven-v4-turbo-to-chase-openai-in-voice-agents/)\n\nThat Jev comparison matters more than the raw score. Decision models, small models that output probabilities instead of text, are a genuinely new slice of AI infrastructure, and until recently Jev had it mostly to itself. Benchmarks from Lyzr found Jev was the best zero-shot router across three separate routing tasks: accuracy between 0.80 and 0.91, a median latency around 350 milliseconds, at a cost Lyzr pegged at two to seven cents per thousand calls.\n\nDrex v2.0 beats Jev head to head in an eight-game arena test, 60.9% to its rival, and edges it out on email classification with an AUC of 0.984 against 0.981. Those are narrow margins. But the fact that two well-funded startups are now racing on the same obscure benchmark, the Decision Index, tells you the category has real demand behind it, not just one company's bet.\n\nNace.AI isn't a fly-by-night entrant chasing a trend. The company raised $21.5 million in a seed round led by Walden Catalyst in May 2026, on top of $5 million raised in March 2025, bringing its total to $26.5 million, according to a report from Pulse2. Yahoo co-founder Jerry Yang and an Intel executive are both named as investors - an unusual amount of star power for a company most people had never heard of a year ago. CEO Dos Bakytzhan previously ran GOAT AI, a consumer app that stayed in Europe's top App Store charts for 37 weeks before Spain's Newry Global Media bought it in 2025. Co-founder Zhanibek Datbayev came from Meta's AI engineering team.\n\nHere's the thing about this moment in agent infrastructure: everyone assumed the fight would be over orchestration, whichever framework gets to sit on top and coordinate the agents underneath. Drex and Jev point somewhere else entirely. They're both betting the real bottleneck is smaller and dumber than that: agents waste enormous compute asking a 400-billion-parameter language model to decide something a lookup table could answer - if only that lookup table could reason over messy, real-world context.\n\nFrankly, that's a more interesting bet than another orchestration layer. Exa's search infrastructure and ElevenLabs' low-latency voice models are both chasing the same instinct, that speed and precision at one specific layer of the agent stack beats being a jack of all trades. Exa raised $250 million at a $2.2 billion valuation in May, a sign investors already believe narrow, fast infrastructure components are where the money sits, not in whoever builds the flashiest end-user agent.\n\nWhether Drex displaces Jev or just forces it to get better is beside the point for most founders watching this space. The real signal is that decision models now have two credible vendors, a shared benchmark, and venture capital on both sides. That's what a category looks like right before it either consolidates around a winner or fragments into a dozen specialized variants: one for routing, one for fraud, one for content moderation. Six months from now, the number of companies playing in this exact lane will tell you which direction it went.\n\n**Also read:** [ElevenLabs launches Eleven v4 Turbo to chase OpenAI in voice agents](https://startupfortune.com/elevenlabs-launches-eleven-v4-turbo-to-chase-openai-in-voice-agents/) • [Jensen Huang Warns Europe Not to Overthink AI Rules or Risk Falling Behind](https://startupfortune.com/jensen-huang-warns-europe-not-to-overthink-ai-rules-or-risk-falling-behind/) • [Why 4Humanity Built a Searchable Record of the Good Already Happening Across America](https://startupfortune.com/why-4humanity-built-a-searchable-record-of-the-good-already-happening-across-america/)\n\n[Skyfall AI unveils SKY-001, a world model built only for engineering](https://startupfortune.com/skyfall-ai-unveils-sky-001-a-world-model-built-only-for-engineering/)\n\nSkyfall AI claims SKY-001 matches GPT-6 Astra's CAD accuracy at 79.7% IoU while running 51 times faster and up to 150 times cheaper, with commercial launch targeted for early 2027. - [specialized world model for engineering design automation](https://startupfortune.com/skyfall-ai-unveils-sky-001-a-world-model-built-only-for-engineering/) - [how world models improve CAD software productivity workflows](https://startupfortune.com/skyfall-ai-unveils-sky-001-a-world-model-built-only-for-engineering/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/nace-ai-s-drex-v2-0-bets-agents-don-t-need-words-to-make-a-decision", "canonical_source": "https://startupfortune.com/naceais-drex-v20-bets-agents-dont-need-words-to-make-a-decision/", "published_at": "2026-10-20 18:11:40+00:00", "updated_at": "2026-10-20 18:19:14.391012+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-startups", "large-language-models", "ai-infrastructure"], "entities": ["Nace.AI", "Drex v2.0", "Drex v1.5", "Qwen3.6-35B-A3B", "TypeSafe AI", "Jev 1.13.0", "Walden Catalyst", "Dos Bakytzhan"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/nace-ai-s-drex-v2-0-bets-agents-don-t-need-words-to-make-a-decision", "markdown": "https://wpnews.pro/news/nace-ai-s-drex-v2-0-bets-agents-don-t-need-words-to-make-a-decision.md", "text": "https://wpnews.pro/news/nace-ai-s-drex-v2-0-bets-agents-don-t-need-words-to-make-a-decision.txt", "jsonld": "https://wpnews.pro/news/nace-ai-s-drex-v2-0-bets-agents-don-t-need-words-to-make-a-decision.jsonld"}}