Most foundation-model releases this year have gone one direction: bigger, more general, cloud-hosted. LFM3, which Liquid AI announced on 27 October, goes the other way. It’s a family of 2B-parameter-class models — one each for text, vision, audio, “omni” (all three at once), and two calibrated decision-making variants — all branching from a single 1.7B shared backbone, and all small enough to fine-tune on one GPU and run on a phone.
The pitch isn’t “a smaller GPT.” It’s a starting point: pick the model for your input type, fine-tune it on your own data, and deploy it where your application actually runs — which for a lot of creative and interactive work means offline, on a device, with no API bill per inference.
What’s actually new versus LFM2
Liquid’s own comparison is specific: LFM3-1.7B outperforms the previous LFM2.5-2.6B while using 35% fewer parameters. The gains come from a redesigned “tri-hybrid” layer mix — 13 gated short convolutions, 5 sliding-window attention layers, 6 global attention layers — tuned via what Liquid calls hardware-in-the-loop architecture search across 14 device and backend combinations, from a Raspberry Pi 5 up to cloud GPUs and the Qualcomm Hexagon NPU. The text model’s context window stretches to 256K tokens, long enough for an on-device agent to hold a genuinely long working context without phoning home.
The vision and audio encoders are new in-house builds, not stock components: the 164M-parameter vision encoder reportedly matches Qwen3.5-2B’s encoder in accuracy while using 2.2x fewer FLOPs, and the video path embeds each pair of sampled frames into one temporal slice to cut sequence length. Every one of the six models fits in roughly 1GB at 4-bit precision.
Why this matters for people building tools, not just chatbots
A 2B vision-language model that fine-tunes on one GPU is a different proposition for a creative-tools builder than a frontier model behind an API: you can specialize it on your own dataset — a particular visual style, a particular sound palette, a particular domain’s jargon — and ship the result inside an app rather than a subscription. The omni model (LFM3-Omni-2B) takes text, images, video and audio together and produces text or speech back, which is the shape most “assistant embedded in a creative tool” projects actually need, rather than a general chatbot wearing a plugin.
The tracker caught this in real time
This site’s own daily Hugging Face tracker recorded the effect directly: in the 24 hours between our 2026-10-29 and 2026-10-30 snapshots, LiquidAI/LFM3-1.7B-GGUF went from 61,499 to 97,184 downloads (+58.0%), and LiquidAI/LFM3-VL-2B-GGUF from 24,617 to 36,468 (+48.1%) — both independently confirmed by re-querying the repos directly. That’s organic adoption curve, not a one-day launch spike, which is the more interesting signal.