Liquid AI Challenges the Parameter Race with LFM2.5-8B-A1B

Liquid AI Challenges the Parameter Race with LFM2.5-8B-A1B

Liquid AI recently unveiled its latest artificial intelligence model, the LFM2.5-8B-A1B, which questions the notion that a model's performance is dependent on its size. In an industry where model size, measured in billions of parameters, often equates to power, Liquid AI is taking a different approach. The goal is to prove that a more compact model can compete with larger rivals without requiring massive infrastructure.

A Model Optimized for the General Public

The LFM2.5-8B-A1B is designed to operate efficiently on consumer devices, without relying on the cloud. Based on a Mixture-of-Experts (MoE) architecture, this model utilizes 8 billion parameters but activates only a portion for each request. This method reduces resource requirements while maintaining high performance.

Liquid AI presents this model as an intelligent personal assistant, capable of performing everyday tasks, using various tools, and following complex instructions. The ambition is to enable this advanced AI to run directly on devices such as laptops or smartphones.

According to Liquid AI, the LFM2.5-8B-A1B competes with larger AIs in instruction-following tests and agentic tasks. Moreover, it stands out for its speed, being the fastest model in its category on CPU and GPU. It is compatible with several popular tools in the AI ecosystem, including llama.cpp, MLX, vLLM, and SGLang.

Notable Advancements

Compared to its predecessor, the LFM2-8B-A1B launched in 2025, this new version brings significant improvements. The context window has been expanded from 32,768 to 128,000 tokens, allowing for the analysis of longer documents and reasoning over more extensive sequences.

The model's vocabulary has also doubled, increasing from 65,536 to 128,000 tokens, which enhances performance for languages using non-Latin writing systems, such as Hindi, Thai, Vietnamese, Indonesian, and Arabic.

While the overall architecture remains similar, combining MoE, GQA, and short-gate convolution blocks, the model has benefited from more ambitious training. The pre-training volume has increased from 12 to 38 trillion tokens, with reinforcement learning phases to improve reasoning and reduce hallucinations.

One of the major innovations is the focus on explicit reasoning. Unlike the previous version, the LFM2.5-8B-A1B generates a chain of thought before producing its final answer, thereby improving the quality of results without compromising performance due to the efficiency of its MoE architecture.

In summary, Liquid AI is not seeking to increase the raw power of its model but to demonstrate that a compact AI can advance without following the frantic race for parameters.