AI BriefingHolo3.1: Local/Mobile Agent Released
Key point
The next-generation computer use agent series Holo3.1, delivering strong performance in local and mobile environments, has been released.
Details
Holo3.1 is a model series that significantly improves robustness across diverse GUI environments and agent frameworks, including web, desktop, and mobile.
Key Features and Improvements:
- Enhanced mobile automation: The 35B-A3B model achieved 79.3% performance on the AndroidWorld benchmark, boosting mobile environment control capabilities.
- Diverse model lineup: From small models of 0.8B, 4B, and 9B to the high-performance 35B-A3B model, supporting cost-efficient and private deployment.
- Optimized for local inference: For the first time, quantized checkpoints such as FP8, Q4 GGUF, and NVFP4 are provided, maximizing local execution speed while minimizing performance degradation.
- Function-calling support: Supports native function-calling protocols, enhancing integration with various agent stacks.
Inference Performance: Using NVIDIA's NVFP4 W4A16 setup, it recorded 1.74x faster token throughput compared to BF16 on DGX Spark, showing an overwhelming speed improvement.
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.