1.2B / 1.5B / 1.6B (family variants with hybrid architecture)
LFM Open License v1.0
Comprehensive on-device model family built on Liquid AI's hybrid LFM architecture and released as a unified LFM2.5 generation. Official release materials describe a scaled training recipe with extended pretraining from 10T to 28T tokens plus expanded post-training and reinforcement learning, targeting strong edge performance under tight memory budgets. The family includes text-generation checkpoints (LFM2.5-1.2B-Base, LFM2.5-1.2B-Instruct, LFM2.5-1.2B-JP), a vision-language checkpoint (LFM2.5-VL-1.6B), and an end-to-end audio-language checkpoint (LFM2.5-Audio-1.5B). Instruct benchmark tables report strong sub-2B results across GPQA, MMLU-Pro, IFEval, IFBench/Multi-IF, AIME25, and BFCLv3; JP release notes emphasize Japanese instruction-following gains; VL notes emphasize improved multi-image/OCR and multilingual visual understanding; Audio notes highlight speech/text I/O with a custom LFM-based detokenizer and lower-latency generation. Deployment support is published across Transformers, llama.cpp (GGUF), MLX, vLLM, and ONNX, with partner optimization notes for AMD and Nexa/Qualcomm NPU paths. Release framing focuses on private, fast, always-on edge intelligence across laptops, mobile devices, and embedded scenarios.