Liquid AI released the open-weight vision-language model LFM2.5-VL-3B on Hugging Face on Wednesday, making it available without a waitlist and capable of running entirely on consumer hardware, including phones, with no image data routed to a server. The 3.1-billion-parameter model matches rivals carrying 50% more parameters in benchmark results, all of which are vendor-reported with no independent third-party replication at time of publication. Its text backbone, LFM2.5-2.6B, uses gated short convolutions and grouped query attention blocks, descending from Liquid Time-Constant networks developed at MIT CSAIL, compressing sequence history into a fixed-size state instead of a growing KV cache. This keeps memory footprint nearly constant regardless of context length, fitting in about 3.3 GB of device memory while decoding 228 tokens per second on an Apple M5 Max and 116 tokens per second on an AMD Ryzen AI Max+ 395. A vision-language pipeline runs on a Galaxy S26 Ultra at 20 tokens per second, fully on-device and offline-capable. The vision encoder is Google's SigLIP2 400M NaFlex, which processes images at their actual proportions rather than square crops.