A New Standard for Agents

Holo3 set a new standard on desktop computer use benchmarks, recording OSWorld-Verified 78.85%.

The model runs with 10B active parameters (122B total), aiming for lower cost compared to large closed-source models. The weights were released on Hugging Face under the Apache 2.0 license, and the API is also available with a free tier.

The core is the agentic learning flywheel.

• Synthetic Navigation Data: Generates scenario-specific navigation examples using human and generated instructions
• Out-of-Domain Augmentation: Programmatically expands scenarios to handle exceptional cases
• Curated Reinforcement Learning: Selects and reinforces samples through filtering and RL

For training, they built a Synthetic Environment Factory that mimics real corporate systems. A coding agent generates websites from scratch based on scenario specifications, and verification scripts confirm whether end-to-end tasks are feasible.

Evaluation was conducted using H Corporate Benchmarks. It consists of a total of 486 multi-step, practical work tasks, with categories as follows:

• E-commerce
• Business software
• Collaboration
• Multi-App

In particular, Multi-App tasks include long-horizon workflows such as reading equipment prices from a PDF, cross-checking against each employee's remaining budget, and then automatically sending approval or rejection emails to each person. Beyond single-app tasks, the design targets the ability to maintain state and intent while moving across multiple systems.

As a result, Holo3 claims to outperform competing models and the Qwen3.5 family on single-application benchmarks, emphasizing that a specialized training pipeline can create a bigger performance gap than sheer parameter count. As a next step, they presented Adaptive Agency, previewing a direction where the model learns to handle new, unfamiliar corporate software in real time.