LiquidAI Releases LFM2.5-24B-A2B On-Device Reasoning Model
LiquidAI has introduced LFM2.5-24B-A2B, a 24-billion parameter reasoning model featuring 2.45 billion active parameters designed for fast, private on-device agentic and tool-use tasks. It represents the largest and final model in the LFM2.5 family, trained on a massive budget of 36 trillion tokens with a context length of 131,072 tokens. This model achieves competitive agentic and tool-use performance compared to significantly larger models while operating at high speeds and low computational costs. Its efficient design enables fast local inference, making powerful reasoning capabilities accessible directly on consumer hardware and edge devices. The architecture utilizes 40 total layers combining 30 double-gated short convolution blocks with 10 GQA layers and a Mixture of Experts structure featuring 64 experts with top-4 routing. It achieves inference speeds up to 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on a single NVIDIA B300 GPU.
## BACKGROUND
Mixture of Experts (MoE) is a machine learning architecture where a large total number of parameters are divided into specialized sub-networks called experts, but only a small subset of these parameters are actively used for each generated token to reduce latency. GGUF is a self-contained file format optimized to store large language models efficiently for community hardware and local execution frameworks.