Announced October 27, 2026, LFM3 is the third generation of Liquid Foundation Models — text, vision, audio, omni, and decision models on one compact backbone, aimed at phones, PCs, and edge devices rather than the datacenter.
2026-10-28 · by Moe Ameen
On October 27, 2026, Liquid AI introduced LFM3, the third generation of its Liquid Foundation Models, under the banner "Every Modality for Specialized Intelligence." Rather than one flagship model, the first release is a family of six open-weight models — all in the 2B-parameter class — branched from a single shared backbone pre-trained on 22 trillion tokens. The lineup 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 whole point is small-and-specialized: one base you can fine-tune into whatever narrow job you need.
Architecturally LFM3 is a hybrid. Across its 24 layers Liquid AI mixes gated short convolutions, sliding-window attention, and a handful of global attention layers, which is how it keeps the memory footprint low while extending context to 256K tokens in a later training stage. The efficiency numbers are the headline: Liquid AI 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 the next model it tested on-device — with a 128K-token key-value cache of about 1.6 GB versus several gigabytes for comparable small models. The company also claims LFM3-1.7B outperforms its own 2.6B predecessor while using about 35% fewer parameters.
On benchmarks, LFM3 leans hard into agentic and instruction-following tasks — tool calling, function calling, and web-automation suites — where Liquid AI reports it beating models several times its size, while conceding that larger models still lead on general knowledge, long-context recall, and agentic coding. Those figures are self-reported and have not yet been independently verified, a caveat worth keeping in mind. The base and post-trained weights are available on Hugging Face, the Liquid Playground, and the Liquid AI API. Coverage of the launch reports the weights ship under an "LFM Open License" that is free for commercial use below roughly a $10M annual-revenue threshold; confirm the exact terms on the model card before you build on it.
Liquid AI framed LFM3 as an on-device and edge story, with launch commentary from AMD (Ryzen AI PCs), Ambarella (edge hardware), and Qualcomm (Snapdragon devices) all pointing at the same idea: multimodal AI that runs locally, privately, and without a server round-trip.
LFM3 is genuinely exciting if you build software — a tiny, private, multimodal model you can run on a laptop is a new primitive. But if you are a creator, "a 2B model you can fine-tune on one GPU" is not a plan; it is a parts bin. Between the model weights and a week of on-brand posts across every platform sits transcription cleanup, vertical reframing, caption styling, carousel layout, blog and newsletter drafting, brand-voice enforcement, scheduling, and publishing — none of which LFM3 does. That gap is the entire job, and it is the job [Kompozy](/) already does as a managed engine, so you do not have to assemble it around an open model.
The practical read for most people: you do not need to wait for LFM3 to be wired into a product, and you do not need to run a GPU, to get the outcome the launch is pointing at — fast, multimodal, on-brand content. Drop a long video or a voice memo into Kompozy and it generates the finished set in [18 formats](/glossary/output-buckets): a [Clipped Short](/glossary/clipped-short) with burned-in [captions](/glossary/caption), a brand-exact [Carousel](/glossary/hyperframes), [Persona Shorts](/glossary/persona-shorts) talking-head video, quote graphics, a blog article, and a newsletter — all 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). If and when a tool you use embeds LFM3 for on-device drafting or captioning, that is upstream of Kompozy: the model makes a raw asset, Kompozy turns it into published content. Either way, the finished pipeline is the part that ships — and that is the part you can run today.
LFM3 is the third generation of Liquid AI's 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 decision/agent models — built from one shared backbone and designed to run efficiently on devices like phones, laptops, and edge hardware rather than in the datacenter.
The base and post-trained weights are open and available on Hugging Face, the Liquid Playground, and the Liquid AI API. Launch coverage reports the weights ship under 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 a commercial product on it.
No. LFM3 is a foundation-model family — it generates text, understands images and video, and handles speech, but it does not caption for vertical feeds, build carousels, write blog posts, or publish to any platform. Turning model output into finished, scheduled posts is a separate job that an engine like Kompozy handles across eight social platforms plus blog and email.
Running locally, it can transcribe audio, read images, and draft text without uploading anything — useful for privacy-sensitive footage and for cutting per-call API costs. In a content workflow that output is a raw ingredient; Kompozy is the layer that reframes, captions, formats, and publishes it into on-brand posts everywhere.