Nace.AI Launches Drex Max Decision Model, Open-Sources Drex V2 and V1.5

In late October, Nace.AI released Drex Max, a multimodal decision model positioned as a "screen-aware decision model" for powering computer use / browser-operating agents, while also open-sourcing its sibling models Drex V2 and V1.5 on Hugging Face. The company says Drex Max was heavily distilled from millions of decision samples and claims top-three placement on nearly all public benchmarks.

Confirmed

• @DotaMate first described Drex Max 2.0 as a 27B-parameter, locally deployable SOTA decision model (which conflicts with later posts saying Drex Max is API-only while V2/V1.5 are open-sourced)
• A series of posts from @nischaytwt revealed: Drex Max takes text, images, charts, and video frames as input and does not generate text; a single forward pass outputs calibrated probabilities for each option (yes/no, multiple choice, ordinal scoring), making it suited for routing, classification, and policy selection
• Performance metrics: single-decision latency under 100ms (another post says <200ms); results within 1 second at 262K token context; priced at $0.06 per million input tokens; already listed on OpenRouter (nace-ai/drex-max)
• Benchmark results: Decision Index 67.59; Decision Index Vision(public) overall score 74.5 (including CV-Bench 86.3), versus 63.9 for GPT-6 Luna Decisions
• Open-source components: Drex V2 is a 35B MoE and Drex V1.5 is 9B, both supporting local/on-device deployment

Unconfirmed

• All posts come from the launch party or reposts; the "top three on nearly all public benchmarks" claim and the comparison with GPT-6 Luna lack independent third-party verification
• The relationship between m1's "27B locally deployable Drex Max 2.0" and the later account of Max being API-only with V2/V1.5 open-sourced is not clarified in the materials

Why it matters

• Decision models compress an LLM's generative output into option-level calibrated probabilities, emphasizing low latency and low cost ($0.06 per million input tokens) — a differentiated technical approach for agent routing and computer use scenarios
• The open-source 9B/35B MoE versions offer a viable option for local and on-device deployment, and the dual API-plus-open-source track covers different usage thresholds

2026-10-21 ~ 2026-10-23 · 6 related posts

• Nace.AI Launches Drex Max 2.0, a 27B Local Decision Model for Computer Use — DotaMate · 2026-10-21
• [source] Drex Max debuts: heavily distilled decision model claims top-3 on public benchmarks at $0.06/M tokens — nischay_twt · 2026-10-23
• [source] Drex Max shares vision benchmarks: 74.5 Public Vision Index vs GPT-6 Luna's 63.9 — nischay_twt · 2026-10-23
• Nace.AI launches Drex Max, a vision-based decision model for computer-using agents — nischay_twt · 2026-10-23
• Nace.AI's decision model Drex Max is now live on OpenRouter — nischay_twt · 2026-10-23
• [source] Drex V2.0 (35B MoE) and V1.5 (9B) open-sourced for on-device deployment — nischay_twt · 2026-10-23