Liquid AI's Tiny LFM2.5-230M Outperforms Larger Models, Runs Anywhere
Liquid AI, founded by former MIT computer scientists, today launched LFM2.5-230M, its smallest AI language model yet. This 230-million-parameter foundation model is explicitly designed for on-device agentic workflows,
Liquid AI, founded by former MIT computer scientists, today launched LFM2.5-230M, its smallest AI language model yet. This 230-million-parameter foundation model is explicitly designed for on-device agentic workflows, capable of running "anywhere" on smartphones, laptops, and robotics. It reportedly outperforms models over four times its size in data extraction, highlighting a shift towards architectural efficiency in AI.
LFM2.5-230M targets developers and engineers building lightweight data extraction pipelines and autonomous edge systems. It operates under a dual-use commercial license, free for individuals and companies under $10 million in annual revenue, but requiring a paid enterprise agreement for larger corporations. This strategy balances grassroots adoption with intellectual property protection.
LFM2.5-230M: Engineered for Edge Performance
The model leverages Liquid AI's LFM2 framework, a hybrid architecture. It combines gated short-range convolutions with grouped-query attention, ensuring efficient information processing and long context management (32K context window) without the high memory demands of traditional transformers.
Its compact memory footprint, under 400MB, allows robust performance on constrained hardware. It achieves a 213 tokens/second decode speed on a Samsung Galaxy S25 Ultra and 42 tokens/second on a Raspberry Pi 5. Its GPU inference stack also delivers lower end-to-end latency compared to competing small models.
Enterprise Value in AI ETL
Traditional rule-based ETL systems are rigid and brittle. LFM2.5-230M facilitates the industry's shift to "AI ETL," where machine learning automates data structuring and adapts to changes. This is critical for automating repetitive tasks like invoice parsing or address formatting.
Using costly, massive cloud models for these routine functions is economically unfeasible. LFM2.5-230M offers a cost-effective, low-latency alternative by running directly on local hardware, reducing reliance on expensive, continuous cloud API calls.
Surpassing Larger Rivals in Key Benchmarks
While other "small" models, such as Weibo's 3-billion-parameter VibeThinker-3B, excel at advanced reasoning, LFM2.5-230M is optimized for data extraction and tool calling. Liquid AI acknowledges it's not for complex math or creative writing.
Despite its microscopic footprint, it excels in its niche. It scored 43.26 on the BFCLv3 tool-use benchmark, surpassing IBM's Granite 4.0-350M (39.58) and Google's Gemma 3 1B IT (16.61). On CaseReportBench for data extraction, it achieved 22.51, beating Alibaba's Qwen3.5-0.8B (Instruct).
Advanced Use Cases and Availability
LFM2.5-230M's strong tool-calling capability makes it an effective skill-selection layer for autonomous systems. Liquid AI demonstrated this on a Unitree G1 humanoid robot, where it translated complex natural language instructions into multi-step plans on-device via NVIDIA's Jetson Orin and SONIC framework.
The base and post-trained models are immediately available on Hugging Face, with native day-one support across inference ecosystems including llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX.
FAQ
Q: What makes Liquid AI's LFM2.5-230M unique compared to other small AI models?
A: Unlike other small models that may aim for broader reasoning tasks with billions of parameters, LFM2.5-230M is a significantly smaller model (230 million parameters) explicitly designed for highly efficient data extraction and tool calling on constrained edge devices. Its unique LFM2 architecture provides superior performance in these specific domains compared to larger models.
Q: Where can the LFM2.5-230M model be deployed?
A: The LFM2.5-230M is designed to run nearly "anywhere" due to its small size and efficient architecture. This includes local deployment on edge devices such as smartphones, laptops, and robotics, enabling on-device agentic workflows without constant cloud connectivity.
Q: What is the licensing model for LFM2.5-230M?
A: The model is available under an LFM Open License v1.0, which functions like open-source software for individuals, researchers, and companies with less than $10 million in annual revenue. However, enterprises exceeding this revenue threshold must enter a separate, paid commercial agreement with Liquid AI for production use.
Related articles
JPMorgan Chase Taps Seattle for Critical AI Control Layer Development
Global financial giant JPMorgan Chase is making a significant strategic investment in Seattle, establishing a new AI software infrastructure team. This pivotal group will build an "AI control layer" to manage the bank's AI operations, aiming to control costs, protect intellectual property, and prevent vendor lock-in.
The Motorola Edge 70 Max is all about power: Android — Key Details
Motorola has launched its new flagship, the Edge 70 Max, designed for power users with a massive 7100mAh silicon-carbon battery and 25W Qi2 wireless charging. It’s the first Android phone since the Pixel 10 Pro XL to support full 25W Qi2, surpassing other Qi2-enabled Androids capped at 15W. The device also offers 90W wired charging and a Snapdragon 8 Gen 5 chip.
DeepMind CEO calls for independent body to regulate frontier AI
DeepMind CEO Demis Hassabis has proposed an independent standards body, modeled after FINRA, to regulate frontier AI models. The body would test advanced AI systems and develop best practices for their release, initially on a voluntary basis before potentially becoming mandatory. This initiative aims to provide technically focused, adaptable oversight to the rapidly evolving field of AI.
OnePlus is reportedly bailing on the US: Oppo — Key Details
OnePlus, and parent company Oppo, are reportedly exiting the US and European markets, with an announcement due shortly. This follows months of rumors and signals a major shift in the Western smartphone landscape.
startups: The web is now mostly bots. Cloudflare is rebuilding its
Cloudflare has launched Precursor, a new defense system, in response to bots now generating over 57% of all web traffic. Precursor monitors entire user sessions to distinguish humans from sophisticated bots, moving beyond traditional single-check methods. This initiative is part of a broader strategy to classify and manage AI agents, control content reuse, and rebuild the web's foundational infrastructure for a machine-dominated internet.
industry: ACRouter picks the smartest AI model per task, beating
A groundbreaking open-source framework, ACRouter, dynamically selects the most capable and cost-effective AI model for any given task, demonstrating a 2.6x cost reduction over Opus-only setups while maintaining performance. It learns and adapts in real-time, addressing limitations of static routing.