Liquid AI runs LFM2 on-device with ExecuTorch
Liquid AI says PyTorch's ExecuTorch was the only inference engine it found that handled LFM2's hybrid caching, which let the company search architectures and ship its open-weight models on phones and laptops.
Liquid AI describes in a blog post how it used PyTorch ExecuTorch to develop and deploy LFM2, its first generation of open-weight models built to run on-device.
When work on LFM2 began in early 2025, the team ran a neural architecture search over many candidate designs. These included different forms of attention, convolutional layers and several kinds of recurrence, and many of them needed hybrid caching to profile token generation realistically. According to Liquid AI, most popular inference engines were tuned for standard transformers and did not support that cache mechanism.
ExecuTorch did support it. The company also pointed to the engine's packaging: a single command converts an exportable graph into one bundle containing the weights, metadata and execution graph. Other engines keep implementation and model information separate, so developers have to manage compatible runtimes for each part.
The switch had costs. Making some of LFM2's custom operations exportable took about two to three weeks of graph reworking. Liquid AI credited ExecuTorch's maintainers with helping solve early problems. LFM2 was running on-device and publicly released by July 2025.
The company tested models from 350M to 4B parameters on an AMD Ryzen AI 9 HX 370 and a Samsung S24. It reported lower prefill and decode latency, a smaller memory footprint and smaller model files. The post does not say what these gains were measured against, and the text gives no specific figures beyond a CPU throughput chart.