Six models in one go. It isn't a marketing stunt, but Liquid AI's strategy to occupy the local inference space, where the marginal cost per token is no longer the only metric that matters. The LFM3 family starts with a 1.7-billion-parameter text model and extends to variants for vision, audio, omni-modal inputs, and two decision models that skip token generation entirely.

The starting point is LFM3-1.7B, an LLM built for local agents and specialized applications. Liquid AI describes it as based on a new "tri-hybrid" architecture: 24 layers combining 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers. It is not a classic transformer, and this architectural choice matters: at 4-bit precision the model occupies about 1 GB of memory, enough to run on any phone or laptop. The context window reaches 256K tokens, a figure that until recently was reserved for much larger models. Liquid AI itself points out that fine-tuning requires a single GPU, a detail that shifts the balance from centralized training to distributed customization.

Alongside the text model, the family includes LFM3-VL-2B for visual understanding, with text, image, and video inputs; LFM3-Audio-2B for audio processing with multilingual speech synthesis across 18 languages; and LFM3-Omni-2B, an end-to-end model that replaces ASR, VLM, LLM, and TTS in a single checkpoint. But the most interesting models for anyone thinking in terms of edge deployment are the two "decision models": d1-Omni-2B and d1-2B. They do not generate tokens, but return calibrated answers in a single forward pass, with declared latencies of 5.8 ms on NVIDIA H100 and 3.7 ms on AMD MI325X for the first, and 5.7 ms and 3.8 ms for the second. When the task is extracting text, numbers, dates, or estimating quantities with an interval, the computational cost collapses because there is no autoregressive generation pipeline.

llama.cpp support has already been merged, removing much of the friction for anyone wanting to self-host on consumer hardware or microservices without relying on proprietary runtimes. This is a structural signal: the AI frontier is not just moving toward larger models in the cloud, but also toward small, specialized models that can live on-premise or on edge devices. For companies with data sovereignty constraints or that want to avoid recurring inference costs, a 1.7B model quantized to 4 bits is a very different object from an LLM with hundreds of billions of parameters: it can be rolled out across device fleets, updated with local fine-tuning, and kept under direct control.

Who gains? Local agent developers, SMEs that want automation without sending sensitive data to third parties, and robotics teams needing low-latency spatial and temporal reasoning. Who loses? Cloud API providers that monetize token consumption on generalist models: if a decision model responds in milliseconds without generating text, the volume of billable tokens drops sharply. Liquid AI's move is not isolated: it belongs to a broader trend toward architectural efficiency and aggressive specialization, where value is measured in milliseconds and gigabytes, not just parameters. One open question remains: will the quality of these 1.7–2.2 billion parameter models be enough for complex tasks? Liquid AI itself advises against using them for agentic coding or multi-step reasoning. But for extracting data from a photo, controlling a device, or answering a question about a video, latency and footprint become the decisive metrics.