Supporting case
Liquid’s local model family adds the downloadable LFM3 release
What happened
Liquid AI released its LFM3 family on October 27. The 1.73-billion-parameter text model supports tool calls and structured responses and is available in formats for local runtimes. Liquid reports roughly one gigabyte of model memory at four-bit precision; that is not a guarantee that a complete application, its context and tools fit in one gigabyte. The model card cautions against complex coding and multi-step reasoning and recommends retrieval for knowledge-heavy tasks. These are publisher claims, not independently measured organisational outcomes.
The weights can be run outside Liquid's hosted service, but the LFM licence has a $10 million annual-revenue condition for commercial use and separate terms beyond it, alongside qualified noncommercial/research provisions. It is not an unrestricted Apache-style release. Licence ↗.
Earlier local components: Liquid AI’s October 7 release makes d1-3B and an experimental smaller multimodal model available as decision components. Rather than generating an extended answer, a component can choose among supplied categories—for example, route a request or classify an image. The release includes downloadable checkpoints and use instructions. Its device-speed demonstrations and limited dataset evaluations do not establish performance across an organization’s full workflow. Release ↗; d1-3B model card ↗; experimental model card ↗.
The weights use the LFM Open License, with a commercial-revenue restriction tied to a $10 million threshold and separate research/nonprofit provisions. They should not be described as unrestricted Apache-licensed models. Actual license ↗.
Small downloadable components offer a local alternative to hosted services, but usable independence also depends on integration, permissions, hardware and licensing. The release is correlated evidence within the existing Liquid family.
Dates and implementation stage
Development: 2026-10-27
