LFM3-VL-2B: a 2B vision model that read our receipt with no GPU
LFM3-VL-2B from Liquid AI is a vision-language model with about 2 billion parameters (1.92B) that reads text and images, built for on-device tasks like reading documents, charts and app screens, and it thinks before answering. It has open weights and is free to use under the LFM Open License unless your company makes $10M+ a year; on Ollama it's a 1.5 to 1.7 GB download. In our test on a CPU-only box with 8 cores and no GPU, we gave it a cafe receipt we made and asked for the total and the priciest item, and it answered $18.20 and the Latte at $4.75, both correct, at about 30 tokens/s using about 2 GB of RAM. Those numbers were measured on our test box and are not typical performance, and one receipt is one test, so check its answers.
Try it
ollama run lfm3-vlInstall Ollama from ollama.com first. The model is a 1.5 to 1.7 GB download. In the chat, type your question followed by the image path, e.g. What is the total? ./receipt.png. We ran it with --verbose --hidethinking added on our CPU-only test box (8 cores, no GPU); speed and memory there are not typical performance.