What if a 1 GB AI model could beat a model more than five times its size?
Not on everything, but on the tasks that matter most for AI agents. That's what Liquid AI is claiming with LFM3, its new family of small, open-weight models released on October 27, 2026.
Here's what LFM3 is, how its architecture works, where it wins, where it doesn't, and who should actually use it.
TL;DR
- LFM3-1.7B is a 1.7B-parameter open-weight model that fits in ~1 GB at 4-bit, with a 256K-token context window.
- It uses a tri-hybrid architecture: 13 convolution layers, 5 sliding-window attention layers and 6 global attention layers.
- It beats Qwen 3.5 9B on agent tasks like WebArena, AutomationBench and function calling.
- It loses to bigger models on agentic coding (SWE-bench) and knowledge (GPQA).
- It's very memory-efficient: 1.6 GB of KV cache at 128K context, versus 4.3 GB for Qwen 3.5 4B.
What Is LFM3?
LFM3 is the third generation of Liquid Foundation Models from Liquid AI. The headline model is LFM3-1.7B, a text model built for local agents and specialised applications.
| Spec | LFM3-1.7B |
|---|---|
| Parameters | 1.7 billion |
| Size at 4-bit | ~1 GB |
| Context window | 256K tokens |
| Architecture | Tri-hybrid (convolution + attention) |
| Pre-training data | 22 trillion tokens |
| Weights | Open-weight |
At 4-bit precision the model fits in about one gigabyte, yet it supports a 256,000-token context window. It also outperforms Liquid's previous LFM2.5-2.6B while using 35% fewer parameters.
Not One Model, a Family of Six
The first release includes six models, all built from LFM3-1.7B-Base and all around 2B parameters:
- Text: for local agents and specialised apps
- Vision: image understanding
- Audio: speech and sound input
- Omni: combines multiple input types
- Two Decision models: specialised from the vision and omni models
Liquid says every model in the family can be fine-tuned on a single GPU.
Why Is It So Fast? The Tri-Hybrid Architecture
Pure transformers use attention in every layer, which gets expensive in memory and speed as context grows. LFM3 mixes three layer types across its 24 layers:
| Layers | Type | What it does |
|---|---|---|
| 13 | Convolution | Fast and lightweight |
| 5 | Sliding-window attention | Looks only at the nearest 256 tokens |
| 6 | Global attention | Full attention across the whole context |
The interesting part is how Liquid found this design. They used hardware-in-the-loop architecture search, benchmarking candidates directly on 14 device and backend combinations. These included phones (Samsung Galaxy S26 Ultra, OnePlus 15), a Raspberry Pi 5, laptops (Apple M4 Max, AMD Ryzen AI), Arm and x86 cloud CPUs, GPUs and a Qualcomm Hexagon NPU.
Where It Wins: LFM3 vs Qwen 3.5 9B
Qwen 3.5 9B is more than five times bigger. Here's where LFM3 comes out ahead:
| Benchmark | What it tests | LFM3-1.7B | Qwen 3.5 9B |
|---|---|---|---|
| WebArena-Verified | Long tasks in a web browser | 38.42 | 28.94 |
| AutomationBench | Tasks inside business apps | 41.29 | 15.10 |
| BFCL V4 | Function / tool calling | 71.39 | 62.83 |
| IFEval | Instruction following | 96.66 | 92.08 |
| LiveCodeBench v6 | Coding problems | 73.46 | 69.86 |
For instruction following, tool calling and agent-style tasks, LFM3 punches far above its weight. On AutomationBench it scores nearly 3x higher than Qwen 3.5 9B.
Where It Loses: The Honest Comparison
Liquid is upfront that larger models still win on agentic coding and knowledge:
| Benchmark | What it tests | LFM3-1.7B | Qwen 3.5 9B |
|---|---|---|---|
| SWE-bench Verified | Fixing real code issues | 43.40 | 51.20 |
| GPQA Diamond | Graduate-level science | 63.37 | 80.40 |
| RULER 128K | Long-context retrieval | 65.56 | 87.11 |
A 1.7B model simply can't store as much world knowledge as a 9B one. LFM3 is a different tool for a different job, not a replacement for big models.
Efficiency: The Real Superpower
KV cache at 128K context:
| Model | KV cache |
|---|---|
| LFM3-1.7B | 1.6 GB |
| Qwen 3.5 4B | 4.3 GB |
| MiniCPM5 2B | 5.6 GB |
On-device speed: measured with llama.cpp at Q4_0 and a 32K context, LFM3-1.7B decodes 1.8x to 2.8x faster than the next fastest model Liquid tested, while using the least memory.
Server scale: on a single NVIDIA B300 with SGLang, it can generate 3.6 billion tokens per day at 1,024 concurrent requests. Liquid also ran 300,000+ LFM3 agents in parallel to analyse the Enron email archive.
Who Should Use LFM3?
Good fit:
- Older laptops or limited RAM
- Local AI agents that rely on tool calling and instruction following
- Phones and edge devices
- Privacy-sensitive work where data must stay on-device
Not a good fit:
- Heavy coding on large codebases
- Tasks that need deep factual knowledge
For those, use a larger model like Qwen, or pair it with LFM3.
AMD, Qualcomm and Ambarella all backed the launch, with plans to run LFM3 on their hardware.
Final Thoughts
| Verdict | |
|---|---|
| ✅ Tiny and fast | ~1 GB, 256K context |
| ✅ Great at agents and instructions | Beats models 5x its size on these tasks |
| ⚠️ Not for heavy coding or deep knowledge | Bigger models still win there |
The direction is what excites me. Capable AI doesn't have to live in a giant data centre. It can run on your laptop.
Next up, I'll install LFM3 locally with Ollama and test it on real tasks. Follow me here or subscribe to The Code Street on YouTube to catch it.
Would you trust a 1 GB model with your daily work? Let me know in the comments. 👇
Sources
- Liquid AI: Introducing LFM3
- Hugging Face: LiquidAI/LFM3-1.7B
All benchmark figures are as published by Liquid AI. Originally published at thecodestreet.dev.
Top comments (0)