LFM2.5-8B-A1B is an edge-native Mixture-of-Experts (MoE) language model developed by Liquid AI. Released in May 2026, it belongs to the LFM2.5 family, which is designed to maximize computational efficiency for on-device deployment. The model contains 8.3 billion total parameters but activates only 1.5 billion per token during inference, enabling it to run locally on consumer hardware like laptops and mobile devices while maintaining performance competitive with much larger dense models.
The model utilizes a hybrid architecture consisting of 24 layers, featuring 18 double-gated LIV (Linear-Implicit-Variable) convolution blocks and 6 Grouped Query Attention (GQA) layers. This design is optimized for high-throughput and low-memory footprints on CPUs, GPUs, and NPUs. It was trained on 38 trillion tokens—a significant increase from the 12 trillion used for the previous LFM2 generation—and incorporates large-scale reinforcement learning to improve instruction following and agentic capabilities.
A primary feature of LFM2.5-8B-A1B is its native reasoning capability; it is a reasoning-tuned model that generates an explicit chain of thought before delivering a final response. It supports a context window of 128,000 tokens, facilitating long-document analysis and complex multi-step tasks. The model also features a doubled vocabulary of 128,000 tokens, which enhances tokenization efficiency for non-Latin scripts, including Arabic, Chinese, Japanese, and Korean.
For optimal performance, the creator recommends using a temperature of 0.2 and a repetition penalty of 1.05. The model is particularly suited for agentic workflows, reliable tool calling, and structured data extraction. It follows a ChatML-like prompt structure for conversation management.