Liquid AI Releases LFM2.5-8B-A1B On-Device MoE Model for Tool Calling
TL;DR. Liquid AI introduced LFM2.5-8B-A1B, an on-device Mixture-of-Experts model designed for tool calling with efficient parameter activation. - The model features 8.3 billion total parameters but activates only 1.5 billion per token for consumer hardware. - LFM2.5-8B-A1B supports a 128K context window and explicitly generates a chain of thought for reasoning tasks. - Training involved a two-stage context extension and reinforcement learning to reduce failure modes and hallucinations.
- LFM2.5-8B-A1B is an on-device Mixture-of-Experts (MoE) model built for tool calling.
- It has 8.3 billion total parameters but activates only 1.5 billion per token, enabling consumer hardware deployment.
- The model includes a 128K context window and performs reasoning by producing an explicit chain of thought.
- Training involved expanding the context window from 32,768 to 128,000 tokens and scaling pretraining volume.
- Reinforcement learning stages were implemented to mitigate 'doom loops' and reduce hallucinations.