Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
AI Engineering
Hacker News·18h ago·5 min read
LFM2.5-24B-A2B: Built for Fast On-Device Agents
TL;DR
LFM2.5-24B-A2B is a new 24 billion parameter reasoning model optimized for fast on-device performance, featuring only 2 billion active parameters. It outperforms similar models in speed and benchmark scores, making it a significant advancement in AI agent technology.
✦ Why It Matters
Explore LFM2.5-24B-A2B for your next AI project to leverage its speed and reasoning capabilities.
Key Takeaways
- Utilize LFM2.5-24B-A2B for faster on-device AI applications.
- Leverage the model's reasoning capabilities for improved answer quality.
- Experiment with the expanded vocabulary for better tokenization.
- Access the model on Hugging Face for local fine-tuning.
Full Summary
LFM2.5-24B-A2B is the latest model in the LFM2.5 family, designed for efficient on-device AI agents with a focus on speed and privacy. It boasts 24 billion total parameters, but only 2 billion are active at any time, allowing it to run up to 1.7 times faster than comparable models.
In benchmark tests, it achieved a score of 48.09 on LiveClaw, outperforming Qwen3.6-35B-A3B, which has 50% more parameters. Key improvements include an extended context window from 32,768 to 131,072 tokens and a vocabulary expansion from 65,000 to 128,000 tokens.
The model employs a reasoning-only approach, generating a clear chain of thought before arriving at answers, which enhances the quality of responses. Both the base and post-trained versions are available for use and fine-tuning on platforms like Hugging Face.