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·May 19·5 min read
OlmoEarth v1.1: A more efficient family of Earth observation models
TL;DR
Processing satellite imagery across large areas requires expensive computation, with sequence length (the number of data tokens a model processes) being a major cost driver that scales quadratically. OlmoEarth v1.1 reduces token sequence length in transformer-based models—neural networks that process sequential data—while maintaining prediction accuracy.
✦ Why It Matters
Engineers can deploy satellite imagery models 3x cheaper while maintaining accuracy, enabling environmental monitoring at national and continental scales.
Key Takeaways
- Evaluate your current model and consider switching to OlmoEarth v1.1.
- Utilize the model's various sizes to fit your compute budget.
- Leverage the efficiency gains for faster satellite imagery processing.
- Explore the technical report for insights on performance changes.
How It Works
OlmoEarth v1.1 reduces compute costs by optimizing the tokenization of satellite imagery. Instead of using multiple tokens for different resolutions, it combines them into fewer tokens, which decreases the overall computational load.
This approach allows the model to maintain performance while significantly lowering the number of tokens processed, leading to faster inference times.