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
Hacker News·14d ago·5 min read
Laguna S 2.1
TL;DR
Laguna S 2.1 is a new 118 billion parameter Mixture-of-Experts model that excels in long-horizon reasoning tasks, achieving competitive results against larger models. It features an impressive context window of up to 1 million tokens and was developed in under nine weeks.
✦ Why It Matters
Explore Laguna S 2.1 for your long-horizon reasoning tasks to leverage its competitive performance.
Key Takeaways
- Utilize Laguna S 2.1 for complex coding tasks.
- Leverage its 1M token context for extensive problem-solving.
- Explore its optimization capabilities for performance improvements.
- Review the available evaluation trajectories for insights.
How It Works
Laguna S 2.1 employs a Mixture-of-Experts architecture, activating 8 billion parameters per token, which allows it to efficiently handle long-horizon tasks. Its training loop has been optimized for longer rollouts, enabling the model to persistently work through complex problems without prematurely declaring success.