Liquid AI introduces the LFM3 model family

The first LFM3 release covers models in the 2B parameter range.[2] Liquid AI describes the family as combining text, vision, audio, omni and decision models on one shared backbone, with the models branched from LFM3-1.7B-Base.[2] The company says both base and post-trained LFM3 models are open-weight and available through Hugging Face, its Playground and its API.[2] For teams assessing adoption, that range offers several model types to evaluate, but the evidence does not specify identical capabilities or deployment requirements across the family.

LFM3-1.7B has 1.73 billion parameters, a 128,000-token vocabulary and a 256K-token context window, according to Liquid AI’s documentation.[3] Liquid AI says its 24 layers comprise 13 convolutions, five SWA layers and six global-attention layers.[2] The company says it pretrained the base model on 22 trillion tokens, with the final 3.5 trillion tokens used in a curated mid-training stage.[2] These figures give technical teams concrete specifications to compare against their own context, memory and infrastructure needs; they do not by themselves establish performance on a team’s workloads.

Liquid AI says LFM3-1.7B outperforms its previous LFM2.5-2.6B model while using 35% fewer parameters.[2] It also says the model matches LFM2-1.2B per parameter on short prompts and runs about 25% faster with 8K tokens in context.[2] These are company-reported comparisons; the evidence does not provide enough detail to treat them as independent evaluations across all use cases.

Liquid AI says LFM3-1.7B uses about 1 GB of memory at 4-bit precision, that every LFM3 model can be fine-tuned on a single GPU, and that the family’s compact footprint suits on-device AI experiences.[2] VentureBeat reported that Liquid AI’s hosted console lists a 128,000-token context limit for the API version, distinct from the 256K-token context specified for LFM3-1.7B in the documentation.[3][4] VentureBeat also reported that businesses exceeding $10 million in annual revenue need a separate agreement for commercial use.[4] Teams should verify the applicable model, API limits and licensing terms before planning deployment.