LFM3: Small but Mighty — The AI Model That Runs on a Raspberry Pi
There's a widespread belief in AI: the bigger the model, the better. More parameters, more capability, more intelligence. GPT-4 has over a trillion. Llama 3 weighs 405 billion. The logic seems simple: size = power.
Then Liquid AI arrives with LFM3 and shatters that logic.
What is LFM3?
LFM3 is the third generation of Liquid Foundation Models, created by Liquid AI, a startup founded by MIT researchers. Their premise is radical: build models that aren't the biggest, but the most efficient per parameter.
LFM3 has 1.7 billion parameters. For context, that's roughly 5 times smaller than Qwen 3.5-9B and 6 times smaller than Gemma 4-8B. Models considered "small" in today's world.
But here's the thing: on many tasks, LFM3 beats them.
David vs. Goliath
The numbers are surprising. On agentic tasks —meaning the model has to use tools, browse the web, manage calendars, make decisions— LFM3-1.7B competes with models several times its size:
- WebArena (real web navigation): LFM3 scored 38.42, Qwen 3.5-9B scored 28.94. A model 5 times smaller wins.
- AutomationBench (task automation): LFM3 scored 41.29, Qwen 3.5-9B scored 15.10. Nearly triple.
- IFEval (instruction following): LFM3 scored 96.66, beating all models in its class.
- BFCL V4 (tool use/function calling): LFM3 scored 71.39, ahead of 8B and 9B models.
It doesn't win everywhere —larger models still dominate in general knowledge and agentic coding— but on the tasks that matter most for real applications (agents, automation, tool use), a 1.7B model holds its own against giants.
1.7B parameters. Runs on a Raspberry Pi 5. And beats 9B models on web navigation and automation. Size isn't everything.
Truly Multimodal
Most small models do one thing: text. LFM3 doesn't. The family includes six models, all derived from the same 1.7B backbone:
- Text — for agents and specialized applications
- Vision — understands images
- Audio — processes sound
- Omni — text + vision + audio together
- Decision (vision) — optimized for decisions from images
- Decision (omni) — decisions from multiple modalities
All this in models of approximately 2B parameters. The idea is simple: instead of using one giant model that does everything mediocrely, you pick the specialized model for your use case, fine-tune it, and run it on your device.
Where Does It Run?
This is where LFM3 gets really interesting. Liquid AI tested the model on 14 hardware combinations:
- Phones: Samsung Galaxy S26 Ultra, OnePlus 15
- Raspberry Pi 5
- Laptops: Apple M4 Max, AMD Ryzen AI
- Cloud CPUs: Arm and x86
- GPUs and Qualcomm Hexagon NPU
On all of them, LFM3 is the fastest model in its class. It decodes 1.8 to 2.8 times faster than the next model. And its memory footprint is tiny: a full 128K-token context needs only 1.6 GB of KV cache, versus 4.3 GB for Qwen 3.5-4B or 5.6 GB for MiniCPM5-2B.
This means you can have an agentic AI assistant running on a Raspberry Pi 5. No data center needed. No cloud subscription. No asking Sam Altman for permission.
Why This Matters
LFM3 isn't just a technical achievement. It's a demonstration that the paradigm is shifting.
For years, the narrative has been: "small models can't do serious tasks, you need the cloud." LFM3 proves otherwise. A 1.7B model can triage your inbox, manage your calendar, research the web, send WhatsApp summaries —all locally, without sending your data anywhere.
And since tokens are free when you run locally, you can have dozens of agents running in parallel on your own machine. No bill. No limits. No dependency.
For SMEs, this changes the equation completely:
- Zero inference cost: you pay for hardware once, not an eternal subscription
- Total privacy: your data never leaves your network
- Customization: you fine-tune the model for your specific domain
- Resilience: it keeps working without internet
The Future Isn't Necessarily Giant
The industry has spent years chasing the biggest model. More parameters, more GPUs, more money, more data centers. But there's an alternative path: small, specialized, efficient models that run wherever you want.
LFM3 won't replace GPT-5 in complex reasoning. That's not what it's trying to do. It's trying to demonstrate that for 90% of real tasks —agents, automation, tool use— you don't need a trillion-parameter model. You need the right model, well-trained, on the right device.
The future of AI isn't just "bigger." It's smarter, smaller, and closer to you.
LFM3 weights are open-weight and available on Hugging Face. You can download them today, fine-tune for your use case, and run on your hardware. No permission needed.
You might also like
Ready for the technology leap?
Don't let your SME fall behind. We implement the AI infrastructure that will give you the competitive edge.
Book Your Free Audit