AI BriefingH Company Unveils High-Throughput Computer Use Model Holotron-12B
Key point
H Company has unveiled Holotron-12B, a computer use agent model that applies a hybrid SSM and attention architecture to maximize inference efficiency.
Details
H Company has released Holotron-12B, a multimodal computer use model developed based on NVIDIA Nemotron-Nano-2 VL. This model goes beyond simple instruction following and is designed as an agentic Policy Model that efficiently performs perception, decision-making, and action within interactive environments.
It maximizes inference efficiency through a hybrid SSM-Attention architecture. To address the quadratic computational cost problem of existing transformers, it combines a State Space Model (SSM) with an attention mechanism. This dramatically reduces memory usage and increases inference throughput when processing long contexts and multiple images.
In WebVoyager benchmark results, it recorded more than 2x the throughput compared to the existing Holo2-8B, achieving a processing speed of 8.9k tokens per second at a maximum concurrency of 100. In addition, WebVoyager performance surged from 35.1% to 80.5%, and it also demonstrated excellent performance on localization and grounding benchmarks such as OS-World-G, GroundUI, and WebClick.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Trajectory Introduces Density-Aware Training to Optimize AI Cost-Per-Task
NASA-IBM Releases Open-Source Multimodal Geospatial AI Model 'Lunar Foundation' for Lunar ExplorationSalesforce Unveils Proprietary AI Model 'Koa' Based on Nvidia Models for Enterprise Workflows