Liquid AI's third-generation Liquid Foundation Models: a family of six open-weight 2B-class multimodal models — text, vision, audio, omni, and agent — built to run on-device.
Last verified · 2026-10-28 · by Moe Ameen
LFM3 is the third generation of Liquid AI's Liquid Foundation Models, announced on October 27, 2026. Instead of one flagship, the first release is a family of six open-weight models, all in the 2B-parameter class, branched from a single backbone pre-trained on 22 trillion tokens. It covers five modalities: a text model (LFM3-1.7B) with a 256K-token context window, a vision model that reads images and video, an audio speech model with speech in and out, an "omni" model that takes text, vision, and audio together and returns both text and speech, and two "decision" models specialized for agentic action. The stated goal is specialized intelligence — a small base you fine-tune into a narrow job, not a do-everything chatbot.
The reason LFM3 is interesting is where it runs. Liquid AI built it to work on-device — phones, laptops, and edge hardware — rather than in a datacenter. It says the text model fits in roughly 1 GB at 4-bit precision, can be fine-tuned on a single GPU, and decodes 1.8 to 2.8 times faster than comparable small models it tested locally, with a much smaller memory footprint for long context. Architecturally it is a hybrid, mixing short convolutions, sliding-window attention, and a few global-attention layers across 24 layers. Launch commentary from AMD, Ambarella, and Qualcomm all pointed at the same use: multimodal AI that runs locally and privately.
On benchmarks, Liquid AI leans into agentic and tool-use tasks — function calling and web automation — where it reports LFM3 beating much larger models, while conceding that bigger models still lead on general knowledge, long-context recall, and agentic coding. Those figures are self-reported and, as of launch, not independently verified. The weights are available on Hugging Face, the Liquid Playground, and the Liquid AI API, under an "LFM Open License" that launch coverage describes as free for commercial use below roughly a $10M annual-revenue threshold — confirm the exact terms on the model card.
The honest boundary for a creator: LFM3 is a model, not a content tool. It can transcribe audio, read an image, draft text, and act as an on-device agent, but it does not caption for vertical feeds, build a carousel, write a publish-ready blog post, or post anything anywhere. Those are a separate workflow.
Think of what LFM3 actually hands a creator as three raw ingredients, not a finished dish: a transcript (from the audio or omni model), a description or set of tags about an image or clip (from the vision model), and a text draft (from the text model). Each is a starting point, generated privately on your own device — and each is exactly the kind of input [Kompozy](/) is built to turn into published content. The clean handoff looks like this: run a voice memo or a long video through an LFM3-powered transcription locally, then drop the audio or video and that transcript into Kompozy and let it do the rest.
From there the mapping is direct. The transcript and the clip become a [Clipped Short](/glossary/clipped-short) with burned-in [captions](/glossary/caption) auto-reframed to 9:16, 1:1, and 16:9; the key points become a brand-exact [Carousel](/glossary/hyperframes) and quote graphics; the draft becomes a blog article and a newsletter; and the same source feeds a [Persona Shorts](/glossary/persona-shorts) talking-head video or a [Persona Frames](/glossary/persona-frames) composite if you want your avatar delivering it. Everything is held to one voice by your [Persona Brief](/glossary/persona-brief), then scheduled and published across the eight social platforms plus blog and email from a single review queue on [Autopilot](/glossary/autopilot). LFM3 keeps the raw processing on your device; Kompozy turns the output into a week of on-brand posts everywhere. You do not need to wire LFM3 into Kompozy — Kompozy works from the files and text, whatever generated them.
LFM3 is Liquid AI's third-generation Liquid Foundation Models, announced October 27, 2026. The first release is a family of six open-weight models in the 2B-parameter class — text, vision, audio, an "omni" model, and two agent/decision models — built from one shared backbone and designed to run efficiently on devices rather than in the datacenter.
That is the point of it. Liquid AI says the text model fits in roughly 1 GB at 4-bit precision and runs fast on-device, with launch partners targeting phones, PCs, and edge hardware. Running locally keeps data private and avoids per-call cloud costs.
No. LFM3 generates and understands content — transcripts, image descriptions, text drafts — but it does not caption for feeds, build carousels, write publish-ready blogs, or post anywhere. Turning its output into finished, scheduled posts is a separate job that Kompozy handles across eight social platforms plus blog and email.
The weights are open and available on Hugging Face, with access also via the Liquid Playground and API. Launch coverage describes an "LFM Open License" that is free for commercial use below roughly a $10M annual-revenue threshold; check the exact license terms on the official model card before building commercially.
Treat LFM3 output as a raw ingredient. Process a source on-device — a transcript, a draft, an image description — then drop the file and text into Kompozy, which turns it into Clipped Shorts, carousels, persona video, a blog, and a newsletter, all on one brand voice, and publishes them across platforms.