Can a 230M AI that runs on a phone really beat models 4x its size?

Liquid AI's LFM2.5-230M, released on June 25, 2026, is a tiny AI with just 230 million parameters (0.23B) that runs entirely on small devices like phones and a Raspberry Pi with no internet, and on some practical tasks like data extraction and instruction following it beats models 3 to 4 times its size. It does 213 tokens per second on a Galaxy S25 Ultra (a token is roughly a word fragment) and 42 tokens per second on a Raspberry Pi 5, but the maker itself does not recommend it for math, coding, or heavy reasoning.

In one line: a 230M-parameter tiny AI runs on your phone with no cloud and beats models 4x its size on specific tasks. It is not an all-purpose assistant; it is a small specialist worker.

What is LFM2.5-230M, and why is it so small?#

Most of today's headline AIs have tens to hundreds of billions of parameters and need a powerful GPU or the cloud. A parameter is a numeric piece of knowledge the model learned during training: more parameters usually means it knows more, but also needs more memory and compute. LFM2.5-230M goes the other way. Its 230 million parameters make it hundreds to thousands of times smaller than today's large models, so it runs on a phone, a laptop, even a Raspberry Pi, without calling the cloud.

Being small buys three concrete benefits. First, your data never leaves the device, so privacy is preserved. Second, it works in airplane mode, giving you true offline operation. Third, there is no cloud fee per call (electricity and hardware aside). The natural next question is: "But isn't a small model just dumb?" Usually yes, but LFM2.5-230M partly overturns that with two tricks.

The first trick is architecture. This is not a plain transformer but an 'LFM2' hybrid. Attention, the heart of a transformer, re-reads the whole sentence for every word, which is accurate but heavy. LFM2 uses that expensive attention only sparingly and handles most of the work with lightweight convolution blocks, which glance at just a few neighboring words. That is far faster and uses less memory, so it stays quick on small hardware.

The second trick is data volume. This small model was trained on a huge 19 trillion tokens. The bowl is small, but it was filled by training very long and very hard to raise its density.

The license is open-weight (the weights are public) but not fully free. Companies above $10M in annual revenue need a separate license for commercial use, while individuals and companies below that can use it for free.

Does a small model really beat one 4x its size?#

The key answer, for specific tasks, is yes. But there is one misconception to clear up: it does not win at everything, only at narrow, specific tasks.

On a data-extraction benchmark that pulls fields out of medical records (CaseReportBench), LFM2.5-230M scores 22.51, beating both Qwen3.5-0.8B (13.83, 3.5x larger) and Gemma 3 1B (2.28, 4.3x larger). On the instruction-following benchmark (IFEval) it scores 71.71, ahead of both (59.94 and 63.49). Keep in mind these are the maker's own published numbers, not yet independently reproduced.

That leads to the second big benefit: it really runs on your device. The model ships day-one with llama.cpp, MLX, ONNX, and GGUF, so Macs use MLX while phones and a Raspberry Pi use GGUF/llama.cpp. And the whole picture is a clear signal of 2026's shift from bigger-is-better toward small, efficient on-device AI.

How do you run it, and what should you watch out for?#

Getting started is simple: grab a GGUF file from Hugging Face and load it into llama.cpp (phone or Raspberry Pi) or MLX (Mac) to run it offline. What it is good at is clear: pulling out structured data, simple tool calls, and following instructions are this model's strong zones.

But there are limits you must know. Liquid AI does not recommend it for math, coding, creative writing, or heavy reasoning. Its broad-knowledge score (MMLU-Pro) is a low 20.25, and it is weak on some agent tool use (τ²-Bench Telecom, 5.26). Small models can also make things up (hallucinate), so add a verification step for important data extraction. Finally, remember that the speed figures above (213 tok/s and so on) are for a 4-bit quantized build with short inputs.

The big picture: Liquid AI demoed the model on the onboard chip (NVIDIA Jetson Orin) of a Unitree G1 humanoid robot, where it acts as the 'edge of the brain' (skill selection) that turns a spoken instruction into a sequence of tool calls. It is a sign that AI is no longer only about giant models: small AIs that run inside the device and do specific jobs well are spreading into everyday gadgets and robots.

Note: the figures here reflect Liquid AI's announcement at the June 25, 2026 release, and most are the maker's own published values not yet independently reproduced. The on-device small-model ecosystem moves fast, so confirm against the latest model card and your own measurements before adopting, and we plan to revisit this quarterly.