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·16h ago·5 min read
Deploy local agents everywhere with LFM2.5-2.6B
TL;DR
LFM2.5-2.6B is a powerful local agent model that excels in tool use and multi-step tasks, rivaling larger models. It achieves efficient inference speeds while maintaining a low memory footprint.
✦ Why It Matters
Consider integrating LFM2.5-2.6B into your projects for enhanced agent capabilities and efficient resource usage.
Key Takeaways
- Utilize LFM2.5-2.6B for high-performance local agents.
- Leverage its strengths in instruction following and tool use.
- Integrate with existing frameworks like llama.cpp and ONNX.
- Run efficient inference on devices with limited resources.
How It Works
LFM2.5-2.6B employs a four-stage training process that includes supervised fine-tuning on agentic data, specialization of teachers for different domains, and multi-domain on-policy distillation. The final stage involves agentic reinforcement learning, where the model learns to interact with various tools and environments through a structured training loop.