We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
Hugging Face Blog·Jun 2·5 min read
Holo3.1: Fast & Local Computer Use Agents
TL;DR
Users faced challenges in running consistent computer-use capabilities across different environments like desktop and mobile. Holo3.1 was developed to enhance robustness in these environments and includes quantized checkpoints for local inference.
✦ Why It Matters
Engineers can leverage Holo3.1 for consistent AI performance across multiple platforms and deployment scenarios.
Key Takeaways
- Utilize Holo3.1 for improved performance across multiple environments.
- Implement quantized checkpoints for faster local inference.
- Consider smaller model sizes for cost-effective deployments.
- Leverage native support for function-calling protocols in third-party stacks.
Full Summary
Holo3.1 builds on the success of Holo3, which was quickly adopted for various workflows like browser automation and business software. Recognizing the need for flexibility, Holo3.1 enhances robustness across environments (web, desktop, mobile), agent frameworks, and deployment targets.
It introduces quantized checkpoints optimized for local inference, including FP8, Q4 GGUF, and NVFP4, enabling faster performance with minimal degradation. For instance, the 35B-A3B model achieves a significant speedup on local hardware, with NVFP4 delivering 1.41 times the throughput of FP8.
Additionally, smaller model sizes (0.8B, 4B, and 9B) are now available for cost-effective and private deployments, making it easier for developers to integrate these agents into their existing systems.