Nace AI Unveils Drex Max: A Multimodal Decision Model for Computer-Use Agents Scoring 67.59
Nace AI has released Drex Max, a multimodal decision model built for agents that operate computers and browsers. Unlike conventional language models, Drex Max reads both text and images and returns calibrated probabilities. It generates no text output. The company reports a score of 67.59 on the public Decision Index 0.3.2, placing it 6.9 points above TypeSafe's Jev 1.14. For agent builders in 2026, this translates into a single, inexpensive, typed call per click.
Overview
Drex Max is designed for the growing class of computer-use agents that need to make reliable, low-latency decisions rather than produce natural-language responses. By returning calibrated probabilities instead of free-form text, the model gives agent frameworks a cleaner signal for action selection, confirmation, and risk assessment.
Key Characteristics
- Size: Parameter count is not disclosed. The API supports up to 262,144 tokens per question, including image tokens.
- Deployment: Hosted API only, available as
drex-maxon the Nace Console. No open weights or quantized builds have been announced. - Performance: 67.59 on Decision Index 0.3.2, the highest score Nace lists for any decision model.
- Strongest Benchmarks: 98.2 on BFCL tool calling and 92.2 on JevBench, compared with 86.7 for Jev.
Why Decision Models Matter for Computer-Use Agents
Computer-use agents increasingly sit at the center of enterprise automation, browser-based workflows, and developer tooling. These agents must decide whether to click, type, scroll, submit, or pause, often under uncertainty and with real consequences. A model that outputs calibrated probabilities rather than text can simplify the decision layer, reduce token overhead, and make agent behavior easier to evaluate and govern.
The 2026 landscape has seen rapid convergence around agentic AI and computer-use systems, with enterprises demanding verifiable, auditable decision paths. Drex Max's typed output fits that requirement: instead of parsing generated text, developers receive a probability they can threshold, log, and test directly.
Benchmarks and Results
The headline figure is the 67.59 score on Decision Index 0.3.2, a public benchmark for decision models. The 6.9-point gap over TypeSafe's Jev 1.14 positions Drex Max as a leading option among available decision models. Notably, the model also performs strongly on tool-calling and decision benchmarks:
- BFCL tool calling: 98.2
- JevBench: 92.2 versus 86.7 for Jev
These results suggest Drex Max is particularly well-suited for agents that must invoke tools and make fine-grained operational choices. However, as with any benchmark, results should be validated against real-world workloads and specific deployment constraints.
What Builders Should Know
For teams building computer-use agents, Drex Max offers a practical integration path through the Nace Console API. The large context window of 262,144 tokens, including image tokens, allows agents to reason over extended visual and textual context, such as full-page screenshots plus accompanying instructions. The lack of open weights means teams should plan for API-based deployment and consider latency, cost, and vendor dependency in their architecture.
The model's design reflects a broader 2026 trend: separating perception and decision from generation. As agentic systems mature, specialized decision models like Drex Max are likely to play an increasingly important role alongside general-purpose multimodal LLMs.
Bottom Line
Drex Max is a focused, benchmark-competitive decision model for computer-use agents. Its combination of multimodal input, calibrated probability output, and strong tool-calling performance makes it a compelling option for agent builders seeking a lightweight, typed decision layer. Whether it becomes a standard depends on real-world validation, pricing, and the continued evolution of the Decision Index and competing models.
via MarkTechPost