Liquid AI has released LFM3, the latest generation of its Liquid Foundation Models, which the company positions as a starting point for specialized AI. The family puts text, vision, audio, omni and decision models on one shared backbone, with each model in the roughly 2 billion parameter range. Developers can choose the model that matches their input type, fine-tune it on their own data on a single GPU, and deploy it wherever their application runs. Liquid AI says its LFM3-1.7B base model outperforms the previous LFM2.5-2.6B while using 35 percent fewer parameters. The gains come from a redesigned architecture found through hardware-in-the-loop search across 14 device and backend combinations, ranging from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and Qualcomm's Hexagon NPU. The base model was pre-trained on 22 trillion tokens, with a final mid-training stage that emphasises agentic traces and tool calls and extends the context window to 256,000 tokens. The company reports that LFM3-1.7B matches LFM2-1.2B on short prompts per parameter and is about 25 percent faster with 8,000 tokens of context. The release includes six models built from the same base: a text model, a vision model that accepts video, an audio model, an omni model that handles text, images, video and audio, and two decision models that return calibrated answers. The text model fits in about 1 GB at 4-bit precision. Liquid AI also introduced an in-house vision encoder of 164 million parameters, which it says matches the accuracy of Qwen3.5-2B's vision encoder while needing 2.2 times fewer FLOPs. The article was cut off before it described the audio encoder in detail. For Australian businesses and professionals, this is mainly a development option rather than an immediate change to cost or obligations. Running small models on local hardware could reduce how much customer data leaves an organisation, which may help with some Privacy Act 1988 considerations, but it does not remove obligations for how personal information is handled. Teams already building AI features may want to test single-GPU fine-tuning on non-sensitive data first. Organisations without active AI development have no need to act on this yet.