LiquidAI ships a 3B vision-language model tuned for edge
Released 13 August, tuned for faster multimodal work on edge hardware. Three billion parameters is the size where the deployment question stops being about cost and starts being about where the data is allowed to go.
A 3-billion-parameter vision-language model from LiquidAI arrived on 13 August, tuned for faster multimodal work on edge hardware. The parameter count is the whole proposition.
Three billion is the threshold where a vision model stops needing a data centre. It runs on a device, which means the image never leaves the device, which means an entire class of regulatory and contractual problems never arises. For anyone processing medical, industrial, or personal imagery, that is worth more than several points of benchmark.
The category is filling out fast. This landed a day after a small multilingual vision-language model from Cohere, and the two together suggest the edge VLM has moved from research curiosity to product segment inside a single quarter.
The evaluation question for small VLMs is different from the one for large ones. Nobody is asking whether a 3B matches a frontier model. The question is whether it clears the bar for one narrow task reliably enough to remove a human check — and that is answered on your own images, not on a leaderboard.
The Open Weights — Open-source AI, tracked daily → · Thunder Compute — Best Open Source LLMs (August 2026) →