Liquid Nanos (www.liquid.ai)

Liquid Nanos is a new family of compact “Liquid Foundation Models” (LFM2) — 350M to ~2.6B parameter checkpoints — engineered to run fully on phones, laptops, and embedded devices. The first public release focuses on task‑specific models (350M–1.2B) for data extraction, English↔Japanese translation, long‑context RAG, tool/function calling, and math reasoning. Liquid reports Nanos achieve frontier‑grade quality on these specialized tasks while occupying 100 MB–2 GB of RAM and evaluated competitively against models hundreds of times larger (e.g., LFM2‑1.2B‑Extract vs Gemma 3 27B; LFM2‑350M‑ENJP‑MT vs GPT‑4o on certain translation benchmarks). All models are on Hugging Face and plug into LEAP and Apollo runtimes. Technically, Nanos combine advanced pretraining with a specialized post‑training recipe (including synthetic data generation for extraction, RL‑based verbosity control for math, and difficulty‑aware reweighting) and targeted fine‑tuning on multi‑turn/multi‑document RAG datasets (1M+ samples across 9 languages). Evaluations use multi‑metric judgments (syntax/format, faithfulness, groundedness, relevance, helpfulness) and bespoke benchmarks to avoid contamination. The result is low‑latency, private, resilient on‑device agents suitable for high‑throughput translation, offline assistants, real‑time tool calling, and scalable RAG pipelines — effectively shifting many use cases away from costly cloud inference toward planet‑scale, cloud‑free deployments.