Holo3-35B-A3B drops, tops OSWorld agent benchmark

Holo3-35B-A3B marks a significant shift in the "Action Model" space by delivering frontier-class GUI reasoning in a highly efficient 3B active parameter package. The model sets a new bar with a 77.8% score on OSWorld-Verified, outperforming significantly larger proprietary models like GPT-5.4 and Claude 4.6 in desktop navigation tasks. Its "Agentic Learning Flywheel" approach, which utilizes synthetic website generation for training, effectively solves common UI grounding failures that plague general-purpose Vision-Language Models. The sparse MoE architecture enables high-speed inference on consumer-grade hardware, making local agentic automation commercially viable. Furthermore, the Apache 2.0 licensing of the weights directly challenges the "agentic" lead currently held by closed-source providers. A strategic pivot to "forward-deployed engineering" under new CEO Gautier Cloix suggests the model is hardened for real-world enterprise edge cases rather than just benchmarks.