Liquid AI Launches LFM2.5-230M: The Model That Democratizes Data Extraction on the Edge
With only 230 million parameters, it outperforms models four times larger in extraction benchmarks and can run on smartphones, laptops, and robots.
June 26, 2026 · 3 min read
TL;DR: Liquid AI launches LFM2.5-230M, a 230 million parameter model that outperforms models four times larger in data extraction and can run locally on devices like smartphones and Raspberry Pi. Its efficient architecture enables 32K token context with under 400MB memory, marking a milestone in edge AI.
What Happened?
Liquid AI, founded by former MIT researchers, has launched LFM2.5-230M, its smallest language model to date. With only 230 million parameters, it is specifically designed for agentic workflows on local devices such as smartphones, laptops, and robotic systems. According to the company, the model outperforms in data extraction benchmarks models four times larger, such as Alibaba's Qwen3.5-0.8B (800M) and Google's Gemma 3 1B (1,000M).
The model uses the LFM2 architecture, a combination of short-range gated convolutions and grouped-query attention that handles a context of up to 32K tokens with memory consumption under 400MB. In tests on a Samsung Galaxy S25 Ultra, it achieves 213 tokens per second in decoding, and on a Raspberry Pi 5, it maintains 42 tokens per second.
Why Is It Important?
This launch marks a milestone in the trend of architectural efficiency versus parametric scaling. While major labs compete over models with hundreds of billions of parameters, Liquid AI demonstrates that competitive performance on specific tasks is possible with a fraction of the computational cost. For businesses, this means being able to run data extraction locally, without relying on cloud connections, reducing latency, costs, and privacy risks.
The model operates under a dual commercial license: free for individuals and companies with annual revenues under $10 million, and requiring an enterprise agreement for large corporations. This facilitates adoption by startups and SMEs while establishing a monetization barrier for scale use.
Consequences for the Market and Users
The emergence of LFM2.5-230M could accelerate the migration of AI workloads to the edge, especially in sectors like manufacturing, logistics, healthcare, and retail, where extracting data from documents, sensors, or telemetry is critical. By offering a viable alternative to pure transformer-based models, Liquid AI pressures competitors like Google (Gemma), Microsoft (Phi-3), and Meta (Llama 3.2) to optimize their small models.
For developers, the model opens the door to autonomous agentic applications on resource-constrained devices, such as offline assistants, service robots, or real-time analysis systems. However, its superior performance in data extraction does not guarantee equal performance in other tasks like complex reasoning or creative text generation, so its use should be evaluated on a case-by-case basis.
What Readers Should Know
- Efficient architecture: LFM2.5-230M uses a mix of convolutions and attention that reduces the quadratic cost of traditional attention, enabling long context with low memory.
- Extraction performance: Outperforms models 4x larger in data extraction benchmarks, but is not designed for general reasoning tasks.
- Ubiquitous deployment: Runs on smartphones, laptops, Raspberry Pi, and robots, with decoding speeds up to 213 tok/s on modern hardware.
- Dual license: Free for revenue <$10M; requires paid license for large enterprises.
- Massive training: Pretrained on 19 trillion tokens, explaining its capability despite the small size.
Context and Comparisons
Liquid AI's move is part of a broader trend toward compact and specialized models. Companies like Microsoft with Phi-3 (3.8B), Google with Gemma 3 (1B), and Apple with on-device models have shown that efficiency is key to mass adoption. However, LFM2.5-230M further lowers the barrier by operating with less than 400MB of memory, something its competitors do not achieve.
The key lies in the LFM2 architecture, which abandons the pure transformer for a hybrid design. This allows Liquid AI to offer fast inference without the memory costs of attention-based models. However, being a proprietary architecture, developers may face integration challenges with standard frameworks like Hugging Face Transformers.
"Liquid AI demonstrates that the future of AI is not just in ever-larger models, but in intelligent architectures that maximize performance per parameter."
Conclusion
LFM2.5-230M is not a one-size-fits-all model, but for data extraction on the edge, it represents a significant advance. Companies looking to automate workflows with strict latency, privacy, or cost requirements should evaluate it. Liquid AI has put forward a real alternative to parametric scaling, and the market will respond.