Liquid AI Unveils LFM2.5-230M, a 230-Million-Parameter Model Optimized for Edge Deployment

Liquid AI, founded by former MIT computer scientists, has released LFM2.5-230M, its smallest language model to date, designed for on-device data extraction and local deployment on smartphones, laptops, and robotics, according to the company’s release blog post. The 230-million-parameter foundation model outperforms larger models like Alibaba’s Qwen3.5-0.8B and Google’s Gemma 3 1B in specific benchmarks, according to the firm.

What is LFM2.5-230M and How Does It Work?

LFM2.5-230M leverages Liquid AI’s LFM2 architecture, a hybrid system that combines gated short-range convolutions with grouped-query attention to process information efficiently. This design allows the model to maintain a memory footprint under 400MB while achieving prefill and decode speeds surpassing comparable models like Gemma 3 1B IT and Granite 4.0-H-350M, according to the company’s internal benchmarks. On a Samsung Galaxy S25 Ultra, it decodes 213 tokens per second, and on a Raspberry Pi 5, it maintains 42 tokens per second, the firm stated.

Why Does It Matter for Enterprises?

Enterprises are increasingly adopting “AI ETL” pipelines to automate data extraction from unstructured sources like PDFs and emails, reducing reliance on rigid, rule-based systems. Liquid AI argues that models like LFM2.5-230M offer a cost-effective alternative to massive cloud-based models such as Claude Opus 4.6, which charges $5.00 per million input tokens. By running locally, LFM2.5-230M minimizes latency and compute costs, according to the company.

How Does It Compare to Other Small Models?

While models like Weibo’s VibeThinker-3B (3 billion parameters) and Google’s Gemma 4 family (2 billion parameters) dominate the “small model” space, LFM2.5-230M operates in a smaller weight class. It scores 43.26 on the BFCLv3 tool-use benchmark, outperforming IBM’s Granite 4.0-350M (39.58) and Google’s Gemma 3 1B IT (16.61), according to Liquid AI’s data. However, the model is not designed for reasoning-heavy tasks like advanced math or creative writing, the company acknowledged.

What Are the Licensing Terms?

LFM2.5-230M is available under the LFM Open License v1.0, which allows free use for individuals and companies with less than $10 million in annual revenue. Entities exceeding this threshold must negotiate a paid enterprise agreement, according to the firm. The license restricts commercial use for larger corporations, protecting Liquid AI’s intellectual property while enabling grassroots adoption.

How Is It Being Applied in Practice?

Liquid AI demonstrated the model’s capabilities by deploying it on a Unitree G1 humanoid robot, where it translated free-form instructions into structured multi-step plans using NVIDIA’s SONIC framework. The model is also available on Hugging Face with native support for frameworks like llama.cpp and ONNX, the company said.

What’s Next for Edge AI?

The release of LFM2.5-230M reflects a broader industry shift toward architectural efficiency over parameter scaling. While major AI firms like OpenAI and Google prioritize trillion-parameter models, Liquid AI’s approach highlights the growing demand for lightweight, edge-friendly solutions. As enterprises seek to reduce cloud dependency, models optimized for local deployment may become critical infrastructure, according to industry analysts.