Liquid AI, founded by former MIT researchers, has released its smallest language model to date, the LFM2.5-230M. With only 230 million parameters, this foundation model is designed to run agentic workflows directly on local devices such as smartphones, laptops, and robots. According to the company, its compact size allows it to run almost anywhere, outperforming models four times its size on specific data extraction benchmarks.
Hybrid architecture that outperforms traditional transformers
The LFM2.5-230M model abandons standard transformer architecture in favor of the LFM2 framework, a hybrid system that combines gated short-range convolutions with grouped-query attention. This design choice enables handling long contexts up to 32K tokens without the quadratic memory costs of pure attention mechanisms. The memory footprint is under 400MB, allowing execution on highly constrained hardware such as the Raspberry Pi 5, where it achieves 42 tokens per second, while on a Samsung Galaxy S25 Ultra with Snapdragon Gen4 CPU it reaches 213 tokens per second.
Sponsored Protocol
Superior data extraction compared to models four times larger
In benchmark tests, LFM2.5-230M scores 43.26 on BFCLv3 for tool calling, dominating IBM's Granite 4.0-350M (39.58) and significantly outperforming Google's Gemma 3 1B IT (16.61). On CaseReportBench for data extraction, it scores 22.51, surpassing Qwen3.5-0.8B (Instruct). Liquid AI emphasizes that this efficiency makes it ideal for AI-based ETL pipelines, where it can extract data from PDFs, emails, or web forms and structure it into JSON without rigid rules.
Dual-use license for developers and large enterprises
The model is distributed under the LFM Open License v1.0, a dual-use commercial license that is not OSI-compliant. For independent developers, researchers, and startups with annual revenue below $10 million, the license is free and allows reproduction, modification, and distribution. Larger enterprises exceeding this threshold must negotiate a separate commercial agreement. This strategy encourages grassroots adoption while protecting Liquid AI's intellectual property.
Sponsored Protocol
The company demonstrated the model's capabilities on a Unitree G1 humanoid robot, where LFM2.5-230M, running entirely on an NVIDIA Jetson Orin module, translated complex commands into multi-step plans, coordinating movements such as walking forward, backward, and kneeling. The model is now available on Hugging Face with native support for llama.cpp, MLX, vLLM, SGLang, and ONNX.
The rise of compact models marks a paradigm shift in artificial intelligence. While giants like OpenAI and Google push toward billion-parameter models, Liquid AI's bet on architectural efficiency shows that high performance can be achieved on edge devices without relying on expensive cloud API calls. For enterprises handling large volumes of unstructured data, such as invoices or telemetry, using massive models would be economically unfeasible. LFM2.5-230M offers a concrete alternative to automate extraction at lower cost and minimal latency, directly on local hardware.
Sponsored Protocol
To explore the topic of AI agents on edge devices, read the article on Notion shutting down Notion Mail to focus on AI agents. Additionally, SpaceX's acquisition of Cursor highlights how trust in AI code generation is the next frontier. For an overview of AI impact in finance, see the article on Google Finance with AI features.
External source: VentureBeat.
