Liquid AI has introduced LFM2.5-24B-A2B, a new open-weight Mixture-of-Experts reasoning model designed specifically for on-device agent tasks. According to reports from MarkTechPost, the model features 24 billion total parameters with just 2 billion active parameters per token, allowing it to run locally on hardware such as the Apple M5 Max where it reportedly achieves a decoding speed of 184 tokens per second.
The Role of Mixture-of-Experts in On-Device AI
Mixture-of-Experts architectures divide a large neural network into smaller sub-networks, or experts, activating only a subset of parameters for any given input. This design allows developers to maintain the high reasoning capabilities of a large model while requiring significantly less compute and memory during inference. By narrowing the active footprint to 2 billion parameters, models like LFM2.5-24B-A2B aim to bridge the gap between heavy cloud-based infrastructure and local execution on consumer hardware, scoring 48.09 on LiveClaw.
Unresolved Technical Details and Future Tracking
Despite the release, the announcement did not specify the exact hardware requirements beyond the Apple M5 Max benchmark, nor did it detail the complete training dataset composition or fine-tuning methodology used by Liquid AI. Observers will need to monitor how the model performs across a wider array of consumer devices and whether future updates will provide broader compatibility documentation.
As edge AI continues to evolve, the practical deployment of efficient reasoning models on local hardware remains a primary area of development to watch.
Lead image: Matheus Bertelli / Pexels
