Designing models around deployment
Liquid AI develops foundation models with attention to memory, latency and the hardware on which inference runs. The company identifies itself as an MIT CSAIL spinout and dates its establishment to 2023. Its work includes model architecture, training and deployment tools, rather than only an API for models built elsewhere. [1]
Its first Liquid Foundation Models appeared on September 30, 2024. The initial language-model series had 1.3-billion, 3.1-billion and 40.3-billion parameters, with hosted testing through playground services. This was access to model outputs, without a public weight release. The company now marks this generation as deprecated. [2]
Downloadable models for local use
LFM2 followed in July 2025 with downloadable 350-million, 700-million and 1.2-billion-parameter checkpoints. Its architecture combined short convolutions with attention. The release targeted local inference, illustrating the company's emphasis on deployment constraints. [3]
LFM2's original license permitted research use and commercial use below a $10 million revenue threshold; larger businesses required a commercial license. Downloadable weights therefore did not mean unrestricted Apache 2.0 terms. The announcement's speed comparisons are Liquid AI's own measurements. [3]
Sources
- Company
Liquid AI · retrieved Oct 23, 2026
- Liquid Foundation Models: Our First Series of Generative AI Models
Liquid AI · Sep 30, 2024
- Introducing LFM2: The Fastest On-Device Foundation Models on the Market
Liquid AI · Jul 10, 2025