LFM3 Liquid AI small models for phones are the most practical small-model release we have seen this year. Our verdict up front: if you need an agent that runs on the device, keeps data local, and costs almost nothing to operate, LFM3 deserves a pilot. If you need a general-purpose reasoning engine or a coding agent, it does not replace a frontier cloud model — Liquid AI’s own model card says so. This review covers what LFM3 is, what each of its six models does, what it costs, where it falls short, and how to test it before committing.
Liquid AI announced LFM3 on 27 October 2026. It is the company’s third generation of Liquid Foundation Models, built around a deliberately unfashionable idea: instead of one ever-larger general model, ship six specialised models of around 2 billion parameters from one shared backbone, designed to run where the data already lives — on phones, laptops, and edge hardware. For a business owner, that framing matters more than any benchmark table: the question becomes how much intelligence you can afford to put inside the product itself.
What LFM3 Liquid AI small models for phones actually are
Every LFM3 model branches from the same base, LFM3-1.7B-Base, pre-trained on 22 trillion tokens. That shared backbone is the point: learn one set of behaviours, and the text, vision, audio, and decision variants behave consistently. The architecture is what Liquid calls tri-hybrid — 24 layers split between 13 gated short convolutions, five sliding-window attention layers, and six global attention layers — mixing cheap local processing with attention where it helps, which is why it can stay small without collapsing on long inputs.
Just as important is how the architecture was chosen. Liquid says it ran a hardware-in-the-loop search across 14 device and backend combinations, including phones, a Raspberry Pi 5, Apple M4 Max and AMD Ryzen AI laptops, Arm and x86 cloud CPUs, GPUs, and the Qualcomm Hexagon NPU. A model searched against real edge hardware is likelier to behave well on your customers’ hardware — and it explains the headline number of this release: the text model fits in about 1 GB of memory at 4-bit precision.
Caution up front: almost every performance number here comes from Liquid AI itself, with independent verification still thin days after launch. Treat benchmarks below as vendor claims to test, not settled facts.
The six models, and which one to pick
LFM3 is not one model. It is a family, and picking the wrong member is the easiest way to be disappointed. Here is what each one does.
- LFM3-1.7B — the text and agent model. With 1.73 billion parameters and a 256,000-token context window, this is the workhorse, post-trained for tool calling and structured output — local agents that sort an inbox, manage a calendar, or send a briefing. It is also the model behind the 1 GB claim.
- LFM3-VL-2B — the vision model. Liquid’s first vision model to accept video input, alongside images, documents, and screens. It answers questions about recorded lectures or phone footage, reads documents and charts, and locates objects. Liquid reports 71.4 across 62 vision benchmarks, up from 57.6 for its predecessor — again, a vendor figure to verify.
- LFM3-Audio-2B — the voice model. This model accepts speech or text and produces text, speech, or both, across 18 languages, covering transcription, synthesis, and translation, with streaming microphone processing and a time to first token as low as 50 milliseconds — the number that makes a voice agent feel responsive.
- LFM3-Omni-2B — the all-in-one model. Liquid’s first omni model combines text, images, video, and audio in a single 2.17-billion-parameter model, answering in text or speech. Use it only when a task genuinely mixes modalities, not as a default.
- d1-2B and d1-Omni-2B — the decision models. These are the most interesting members for business automation, and the easiest to misunderstand. Instead of generating a long reply token by token, they return a typed answer with probabilities in a single forward pass: classification, routing, and “pick the next step” decisions. They are fast and cheap precisely because they do not chat.
Match the model to the job: text work gets LFM3-1.7B, documents get the vision model, voice gets the audio model, routing gets a d1 model.
Pricing: free weights, a revenue line, and a hosted API
LFM3’s pricing is simple, with one clause every business must read before shipping. The weights are open-weight under the LFM Open License v1.0, and Liquid’s pricing page states the policy plainly: free to download, run, and fine-tune, including commercially, for companies with less than $10 million in annual revenue. Research, education, and non-profit use is free with no revenue limit, and fine-tunes stay private — no copyleft forces you to open-source your adapted model.
Above $10 million in annual revenue, commercial use requires a paid enterprise licence from Liquid AI. That threshold is the single most important commercial fact in this review. For a startup, a small agency, or an internal pilot, LFM3 is effectively free; for a company at or above the threshold, it is a model you can test freely but cannot ship commercially without talking to Liquid first. Settle that before you build a product around it, not after launch.
For teams that prefer not to self-host, Liquid offers an API. Its pricing page lists LFM3-1.7B at $0.08 per million input tokens, $0.02 per million cached input tokens, and $0.20 per million output tokens, with the decision model d1 at $0.02 per million input tokens and no output charge — sensible, since a decision model returns a structured answer rather than generated text. Those rates are low, but the real comparison is against running the weights yourself: once the model lives on hardware you already own, the marginal cost of an inference is electricity and maintenance, not a per-token bill.
Who LFM3 is for — and who should wait
LFM3 fits best in four situations. First, privacy-sensitive work — customer conversations, internal documents, health or financial data — where sending data to a cloud API creates a compliance conversation you would rather avoid. Second, offline or flaky-connectivity settings: field teams, warehouses, clinics, and retail floors where a cloud round-trip is a point of failure. Third, high-volume, narrow tasks — triage, extraction, routing, summarisation — where a small model running thousands of times a day beats a large model billed per token. Fourth, cost-sensitive products, where a per-user API bill would destroy the margin.
It is weaker for open-ended knowledge work, long creative writing, complex analysis, and agentic coding — Liquid’s own model card advises against the last two for LFM3-1.7B. It is also weaker for teams that want a managed platform with service-level guarantees and no infrastructure to maintain. Open weights give you control and the operating burden; if nobody on your team can own deployment, monitoring, and updates, the hosted API or a managed cloud model is the honest choice.
Watch-outs and honest limitations
- Vendor benchmarks, not independent ones. The headline comparison — LFM3-1.7B averaging 49.37 across 29 benchmarks against 47.24 for Alibaba’s Qwen3.5-9B, a model more than five times its size — comes from Liquid. Run your own task set before you quote it to a client.
- Long context costs memory. The 1 GB figure is the model at 4-bit precision; Liquid notes a full 128,000-token KV cache needs a further 1.6 GB. Budget for working memory, not just weights.
- Language coverage is uneven. The text models cover 16 languages — reportedly only seven of the EU’s 24 official languages for text, while audio covers more for speech. Check your specific languages against the model card, not the headline count.
- Early tooling rough edges. Launch-week reports already mention MLX builds failing to start for some users. Pilot on your exact target devices and backends — llama.cpp, MLX, vLLM, Ollama, and ONNX are all listed as supported, and they will not all behave identically on day one.
- The licence threshold follows growth. A product free to ship today can require a paid licence after a good year. Build that into your pricing and contracts now, while it is cheap.
How LFM3 compares to the alternatives
Against other small open models, LFM3’s case rests on specialisation, not size. Qwen’s small models and Google’s Gemma family are the natural comparisons: capable, widely deployed, and backed by large ecosystems. Liquid claims its tri-hybrid, hardware-tuned design delivers more agent capability per parameter — fewer tokens, faster on-device decoding (1.8 to 2.8 times faster than the next-fastest compared model on its tests), and one backbone covering vision, audio, and decisions. Test it with a side-by-side pilot on the same tasks, measuring success rate, latency, and cost per task.
Against frontier cloud models, the comparison is not about intelligence at all. A large cloud model will beat LFM3 on breadth, reasoning depth, and knowledge; LFM3 wins on privacy, latency, offline operation, and marginal cost. The right architecture for many products is therefore hybrid: LFM3 on-device for routine, high-volume, privacy-sensitive work, with escalation to a larger model for the genuinely hard cases. Put the cheap model where it is good enough, and spend expensive intelligence only where it changes the outcome — the pattern our AI integration and automation services follow.
From our experience: how we would pilot it
After nine years and 300+ projects, our rule for a launch-week model is the same: a two-week pilot on one real workflow, with success criteria written down first. For LFM3: pick a narrow, high-volume task with a measurable outcome — ticket triage, document extraction, or a voice-command flow — not “general assistance.” Run LFM3-1.7B (or a d1 model) against your current approach on the same 200 examples, scoring accuracy, latency, and cost per task, on the actual target device. Check failure modes, not just averages, and confirm the licence position in writing before the pilot becomes a product.
The usual gotcha with small models is scope creep, not capability. A pilot that succeeds at one narrow task gets quietly expanded — to a second task, then general chat, then coding help — until the model is asked to do a job it was never designed for. Guard the envelope, and LFM3 is a genuinely useful building block. If you want help scoping that pilot, talk to our team, or browse the AI Tools Directory.
Frequently asked questions
Can LFM3 really run on a phone?
Yes, with a qualification. Liquid says LFM3-1.7B takes about 1 GB at 4-bit precision, and its architecture search covered real phones — but a long-context session needs additional working memory, so test on your target handset.
Is LFM3 free for commercial use?
For companies under $10 million in annual revenue, yes, under the LFM Open License v1.0, including commercial use and private fine-tunes. Above that threshold, a paid enterprise licence is required; research, education, and non-profit use is free with no limit.
Which LFM3 model should I start with?
Start with LFM3-1.7B for text agents, LFM3-VL-2B for document and screen understanding, LFM3-Audio-2B for voice, and a d1 model for routing and classification. Reach for omni only when a task genuinely mixes modalities.
What can’t LFM3 do well?
Liquid’s own model card advises against agentic coding and complex multi-step reasoning for LFM3-1.7B, and recommends retrieval for knowledge-heavy tasks. Treat it as a specialised on-device worker, not a general-purpose brain.
How does LFM3 compare with a cloud frontier model?
A frontier cloud model wins on breadth and reasoning depth; LFM3 wins on privacy, latency, offline operation, and marginal cost. The best design for most products is hybrid: LFM3 on-device for routine work, escalating only where extra intelligence changes the outcome.
The bottom line: LFM3 Liquid AI small models for phones are a credible way to put real agent capability inside a product without a cloud bill per action. Pilot one narrow workflow, measure it honestly, confirm the licence for your revenue band, and you will know within two weeks whether on-device AI earns a place in your stack.