Liquid AI LFM3: Six Models That Run On Device
Six roughly 2B models aimed at phones and laptops. The memory number Liquid published is the one worth copying into your notes.
Liquid AI LFM3 arrived on 27 October 2026: six models of roughly 2 billion parameters each, all branched from one base and all aimed at running on the device in someone's hand rather than in a data centre. The release covers text, vision, audio, an omni model that takes several input types, and two decision models specialised from the vision and omni variants.
Six small models is not the interesting part. Small models land every week. The interesting part is a memory number Liquid published that most releases leave out.
What is in the Liquid AI LFM3 release
LFM3-1.7B, the text model, at about 1.73 billion parameters.
Vision, audio and omni models, each described as around 2 billion parameters.
Two decision models, specialised from the vision and omni models.
The text model is a 24 layer hybrid: gated short convolutions, sliding window attention, and global attention together. It carries a 128,000 token vocabulary, supports 16 languages, and takes a context window of 256,000 tokens across input and output combined.
The number that decides whether a model runs on a phone
Parameter count tells you what the weights cost to store. LFM3-1.7B fits in roughly 1 gigabyte at 4-bit precision, which is unremarkable for its size.
What parameter count does not tell you is what the conversation costs as it grows. That is the key-value cache, and it scales with how much context you have actually filled. Liquid published theirs at full context, which is the comparison that matters:
LFM3-1.7B at a full 128,000 tokens: 1.6 GB of KV cache.
Qwen3.5-4B at the same context: 4.3 GB.
MiniCPM5-2B at the same context: 5.6 GB.
On a phone or a laptop with a fixed memory budget, the cache is usually what runs you out of room, not the weights. A model whose weights fit in 1 GB but whose cache wants 5 GB at long context is not an on-device model in any useful sense. This is the figure to ask for when a vendor claims their model runs locally.
Speed, as reported
Using llama.cpp at Q4_0 and a 32,000 token context, Liquid reports LFM3-1.7B as the fastest model on all three devices they tested, decoding 1.8 to 2.8 times faster than the next model, while using the least memory.
At 8,000 tokens of context it is about 25 percent faster than LFM2-1.2B.
It outperforms LFM2.5-2.6B while using 35 percent fewer parameters.
On a single NVIDIA B300 with SGLang in BF16, Liquid reports 3.6 billion tokens per day at 1,024 concurrent requests.
All of these are vendor benchmarks run by Liquid on hardware they chose.
What Liquid says not to use it for
The model card is unusually direct: it says the model is not recommended for agentic coding or complex multi-step reasoning, because of its size. Vendors do not often print the limits of their own release, and taking them at their word saves you a week of disappointment. A 2 billion parameter model is not your coding agent.
Where a model this small actually earns its place
Transcription, classification and extraction that run offline, with no per-token bill and no network round trip.
Anything where the data should not leave the device at all, which removes a whole category of privacy question rather than answering it.
Features with a hard latency floor, where a local answer in 50 milliseconds beats a better answer in 800.
A first pass that filters or routes, handing only the hard cases to a larger hosted model.
The licence is open weights, not open source
The model card names the licence as LFM Open License v1.0 and the weights are on Hugging Face. That is open weights, which is a different thing from open source, and the distinction decides what you are allowed to ship. We laid out the difference in open weights versus open source AI. Read the licence text before you build a product on it, rather than assuming a permissive default.
Liquid's announcement is on the Liquid AI blog, and the text model card with the architecture and licence details is on Hugging Face. For deciding whether any of this warrants changing your stack, see choosing a model without chasing every release, and for filtering the weekly flow of releases like this one, our guide to following AI news.
How did this land?
About the author

Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.


