Holo4 Overview

Holo4 introduces two model sizes: a 27B dense model and a 35B-A3B Mixture of Experts, significantly enhancing capabilities for generalist computer-use agents.

The OSWorld 2.0 scores for Holo4 are 61.7% for the 27B model and 30.9% for the 35B-A3B model, compared to Opus 5.5's score of 81.8%. This indicates that while Holo4 shows promise, it still trails behind its competitors in performance metrics.

Holo4's Agentic Task Factory has produced approximately 10,000 tasks, leveraging its ability to create interactive environments and verifiable tasks from documentation alone. This feature is expected to streamline the development process for AI engineers and ML researchers.

In terms of performance, Holo4's 27B model executed 84 calls with a total of 1.3 million tokens, while its 35B-A3B counterpart's performance remains to be fully assessed. In specific tasks, such as 3D modeling a logo, Holo4 required 94 calls and 1.5 million tokens, whereas the Qwen3.8 model needed 118 calls and 1.9 million tokens.

For game design tasks like creating Pac-Man, Holo4's 27B model made 68 calls with 2.4 million tokens, significantly fewer than Qwen3.8, which required 197 calls and 11.4 million tokens.

Holo4 builds on the previous Holotron 3 model and is priced according to H Models API rates for single runs. The development of Holo4 is part of the ongoing efforts by the NVIDIA Nemotron Coalition, which aims to enhance AI model capabilities.

This follows the release of Holo3.1 on June 2, 2026, and indicates a continued evolution in the Agentic model series, with Holo4 positioned to improve interactions with software through any available interface.