Liquid AI launches Nanos: task-specific 350M-2.6B models that run on-device

JosephJacks_ · x · 2026-09-04

Liquid AI released Liquid Nanos, a family of 350M-2.6B parameter task-specific foundation models (LFM2 family) that claim frontier-grade performance on specialized agentic tasks while running fully on-device with a 100MB-2GB RAM footprint.

First releases include LFM2-Extract (350M/1.2B, multilingual structured data extraction like invoice emails to JSON), LFM2-350M-ENJP-MT (bidirectional EN↔JP translation), LFM2-1.2B-RAG (long-context retrieval QA), and LFM2-1.2B-Tool (tool/function calling for agent workflows).

The pitch flips the deployment model: instead of shipping data to the cloud, compact capable models ship to the device for speed, privacy, and cloud-free economics. Community commentary adds that tiny 0.5B-1.2B models trained for a single task are taking off because they're dirt cheap to run, and 1,000-5,000 high-quality examples often suffice to train a specialized 1B model to frontier-level performance on that task — something local-AI enthusiasts can run and train at home.

Original post →

More from Infra

Infra channel →