On-Device Model
also called Edge model, Local model, Small language model
A small model that runs where the user already is, with no round trip to a cloud API.
A model small and efficient enough to run on a phone, laptop or single-board computer rather than in a data centre, so inference works offline, data stays on the device, and the marginal cost per request is near zero. Liquid AI’s LFM3 family, released October 27, 2026, is the current example: six 2B-class open-weight models on one backbone, with the 1.73B text model fitting in about 1 GB at 4-bit precision.
The trade is capability for placement. A 2B model will not match a frontier model on broad knowledge or complex reasoning — Liquid’s own model card advises against using LFM3-1.7B for agentic coding or complex multi-step reasoning — but for a narrow, high-volume, well-defined job it wins on cost, latency, privacy and reliability at the same time, and it can be fine-tuned on a single GPU.
Check the licence before shipping, not after. The LFM Open License v1.0 is free for commercial use below $10 million in annual revenue and requires a separate agreement above it, and every benchmark figure in the launch is Liquid’s own, not independently verified.
The family structure is the design lesson, not any single model. All six LFM3 models branch from one 1.7B base pretrained on 22 trillion tokens, then specialise by modality: text, vision, audio, omni, and two decision models that return typed answers with calibrated probabilities in a single forward pass instead of generating text. Pick the modality you need rather than renting one generalist and hoping it behaves on your task.
Ten thousand routine decisions a day
A support operation needs every incoming ticket routed and tagged, and the ticket text cannot leave its own systems.
A decision model running on existing hardware answers in milliseconds with a calibrated probability, and only low-confidence cases escalate to a frontier model. The on-device model handles the bulk; the cloud model handles the exceptions. That cascade is where the economics land, not in replacing the frontier model outright.
Assuming small means safe or cheap to ship. You still own updates, evals and the licence terms, and a small model hallucinates just as confidently as a large one.
- blogIntroducing LFM3: Every Modality for Specialized Intelligence ↗
Liquid’s launch post: six 2B-class open-weight models on one backbone pretrained on 22T tokens, with the text model fitting in about 1 GB at 4-bit precision. All benchmark figures are the vendor’s own, not independently verified.
Quantization
Storing weights at lower precision to make a model cheaper to run.
Latency
Total time from request to finished answer.
Fine-tuning
Extra training on your own examples to change how the model behaves.
Model Provider
The company or platform you call to run inference.