Liquid AI Ships an 8B Mixture-of-Experts Model Built for Laptops and Phones

Liquid AI released LFM2.5-8B-A1B, a mixture-of-experts (MoE) language model with 8 billion total parameters and about 1 billion active parameters per token, according to a post from the AI ML Big Data channel. The stated design goal is on-device inference on laptops and smartphones, with chain-of-thought reasoning enabled by default. The release continues the company's LFM2 line.

The architecture is the point of interest. A sparse mixture-of-experts model with a small active-parameter count keeps per-token computation and memory bandwidth low, which is what matters for local inference, while the larger total parameter pool preserves capacity. An active count near 1 billion places the per-token cost within the range that recent mobile neural processing units and laptop accelerators can sustain, which is the practical threshold for usable local latency.

These are vendor figures from a launch announcement, not independent measurements. The claims to verify are the quantized memory footprint on real consumer hardware, tokens per second under realistic context lengths, and whether the default reasoning mode holds up on standard math and code benchmarks against similarly sized dense models. Until third parties publish numbers, the stated parameter counts describe the design, not delivered quality.