Holo3-35B-A3B drops, tops OSWorld agent benchmark
H Company (H) releases Holo3-35B-A3B, an open-weight sparse MoE Vision-Language Model optimized for autonomous computer use. Achieving a state-of-the-art 77.8% on the OSWorld-Verified benchmark, the 35B model (3B active parameters) is engineered for high-speed, multi-step enterprise automation across web and desktop interfaces.
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.
DISCOVERED
11d ago
2026-04-01
PUBLISHED
11d ago
2026-03-31
RELEVANCE
AUTHOR
External_Mood4719