Liquid AI LFM3 is a new family of six open-weight AI models, released on October 27, 2026, and the smallest one needs only about 1GB of memory at 4-bit precision. That means a real text and tool-use model can run on your phone or laptop with no cloud connection. It is not a replacement for the big chatbots, but it is a serious step for private, offline AI.
Here are the numbers worth knowing before you read on.
- Released October 27, 2026, with weights on Hugging Face and a hosted API.
- Six models: text, vision, audio, omni, and two “decision” models.
- The text model, LFM3-1.7B, fits in about 1GB at 4-bit.
- Businesses earning over $10 million a year need a separate license.
If you only care about the catch: these models are small, and Liquid itself says they are not meant for hard coding or long multi-step reasoning.
What is actually in the Liquid AI LFM3 family?
All six models share one small foundation, pretrained on 22 trillion tokens. Liquid calls the design “tri-hybrid” because it mixes convolution layers with attention, which helps keep memory use low. The table shows what each model does.
| Model | What it does | Size |
|---|---|---|
| LFM3-1.7B | Text, tool use, agent tasks | 1.7B |
| LFM3-VL-2B | Reads images and video, answers in text | About 2B |
| LFM3-Audio-2B | Speech and text in and out, 18 languages | About 2B |
| LFM3-Omni-2B | Vision, audio, and text; speech or text output | 2.2B |
| d1-2B | Turns text and images into structured decisions | About 2B |
| d1-Omni-2B | Same, but audio is included too | About 2B |
The text model offers a 256K-token context window in the downloadable version, while the hosted API lists 128,000 tokens. That is plenty for long documents, though memory use grows with context.
How small is “fits in 1GB” in real life?
The 1GB figure covers the model weights only. A running app also needs room for the context cache, which Liquid puts at 1.6GB for a 128K-token context. So plan on more than 1GB once you actually use it.
Liquid also measured full runs with llama.cpp, using a 32K input and Q4_0 files. According to VentureBeat, these were the results.
| Device | Memory used | Speed (tokens/sec) |
|---|---|---|
| Apple M5 Max | 2.25GB | 237.1 |
| AMD Ryzen AI Max+ 395 | 2.75GB | 129.6 |
| iPhone 17 Pro | 1.90GB | 43.6 |
Even 43.6 tokens per second on a phone is faster than most people read. These are Liquid’s own measurements, so treat them as best-case until independent tests appear.
Where it beats bigger models, and where it falls short
Liquid’s benchmarks put LFM3-1.7B ahead of the much larger Qwen3.5-9B on browser and tool tasks. It scored 38.42 on WebArena-Verified versus 28.94, and 71.39 on BFCL V4 versus 62.83. These are company-reported scores, not independent ones.
The picture flips on harder work. Qwen3.5-9B leads on GPQA Diamond (80.40 vs 63.37), SWE-bench Verified (51.20 vs 43.40), and long-context RULER 128K (87.11 vs 65.56).
Liquid is upfront about this. Its docs say the 1.7B model “is not recommended for agentic coding or complex multi-step reasoning due to its small size.” Think of it as a fast helper for narrow jobs, not a brain for everything.
The vision model looks solid for its size. LFM3-VL-2B averages 71.4 across 62 evaluations, against 64.0 for Qwen3.5-2B. Fine-tuning helps too, though not always evenly: in one wind-turbine test, classification accuracy jumped from under 10% to 67%, but detection F1 only reached 0.11.
The license catch for businesses
The models use the LFM Open License v1.0, which is open for most people but not for all uses. Companies with more than $10 million in annual revenue need a separate commercial agreement. Organizations may keep their own adaptations private.
That is different from fully permissive licenses, so check it before you build a product on top. If you want another view of free local models and what their benchmarks really show, our look at Nanbeige4.3-33B covers the same trade-offs.
How to try Liquid AI LFM3 today
You have three easy routes. The quickest is the browser playground at playground.liquid.ai, which needs no install. The weights are in the LiquidAI collection on Hugging Face, and the hosted API lives at console.liquid.ai.
API prices are low. LFM3-1.7B costs $0.08 per million input tokens and $0.20 per million output tokens, and Liquid says it keeps zero data retention. Pricing needs a console sign-in, so contract rates may differ.
Local AI is also moving into browsers. If that interests you, see how Opera WebAI plans to run supported models on your own device.
Frequently Asked Questions
Is Liquid AI LFM3 free to use?
The weights are free to download under the LFM Open License v1.0. Companies above $10 million in annual revenue need a separate commercial agreement.
Can LFM3 run on my phone?
Yes. Liquid measured the text model on an iPhone 17 Pro at 1.90GB of memory and 43.6 tokens per second. Older or lower-memory phones were not reported.
Which LFM3 model should I start with?
Start with LFM3-1.7B for text and tool use. Pick LFM3-VL-2B for images, or LFM3-Audio-2B for speech.
Is LFM3 good at coding?
Not really. Liquid says larger models still win on agentic coding and knowledge, and it advises against using the 1.7B model for complex coding.
Our take
LFM3 is worth a weekend test if you build apps, care about privacy, or want an assistant that works offline. Use it for narrow, well-defined jobs like reading a screen, calling a tool, or transcribing speech. For tough coding or deep reasoning, stay with a larger model, and wait for independent benchmarks before trusting any of these scores.
Leave a Reply