LFM3 is the new family of six open-weight models from Liquid AI, all around 2B parameters and derived from the same 1.7B base model. The text model takes up about 1 GB in 4-bit: it fits on a laptop CPU, on a Raspberry Pi 5, or on a phone. Liquid AI published the weights on Hugging Face on October 27, 2026, with GGUF, MLX, and ONNX versions ready to use, so you can already run it locally with Ollama, llama.cpp, or vLLM.
What is LFM3 from Liquid AI?
LFM3 is the third generation of the Liquid Foundation Models: a single backbone, LFM3-1.7B-Base, specialized into six models covering text, vision, audio, and decisions. It’s a family of SLMs (small language models) designed for on-device inference: the same architecture and efficiency work are reused across all modalities instead of designing each model separately.
| Model | Input → Output | Size |
|---|---|---|
LFM3-1.7B |
Text → text | 1.73B |
LFM3-VL-2B |
Text, image, video → text | 1.9B |
LFM3-Audio-2B |
Text, audio → text, speech | ~2B |
LFM3-Omni-2B |
Text, image, video, audio → text, speech | 2.2B |
d1-2B |
Text, image, video → calibrated responses | ~2B |
d1-omni-2B |
Text, image, video, audio → calibrated responses | ~2B |
The two d1 models deserve a note: they don’t generate free text, but rather calibrated responses. It’s the same category of “decision models” we analyzed in Microsoft-Decision-1 vs Jev. LFM3-Omni-2B is Liquid AI’s first omni model, and LFM3-VL-2B is their first vision model that accepts video.
The rest of the article focuses on LFM3-1.7B, the text model, which is what most developers will try first.
Why does LFM3 take up so little memory?
Because most of the network never looks at the full context. The base model has 24 layers of three types: 13 short gated convolutions, 5 attention layers with a sliding window of 256 tokens, and only 6 global attention layers. That’s what keeps memory consumption low as the context grows.
Other data from the model card:
- Context window: 256K tokens (input plus output).
- Vocabulary: 128,000 tokens; 16 languages, including Spanish and Portuguese.
- Training: 22T tokens of pretraining, with the last 3.5T focused on agentic traces and tool calls, followed by four post-training stages that include agentic RL within real agent harnesses.
- Hardware-in-the-loop design: according to Liquid AI, the architecture was explored by testing it across 14 device and backend combinations: phones, a Raspberry Pi 5, laptops, CPUs, GPUs, and the Qualcomm Hexagon NPU.
How much RAM does LFM3 need?
Around 1 GB for the weights in 4-bit, plus whatever your context window costs. Here are the GGUF file sizes:
| Quantization | Size |
|---|---|
| Q4_0 (QAD) | 1.05 GB |
| Q4_K_M | 1.1 GB |
| Q5_K_M | 1.26 GB |
| Q6_K | 1.43 GB |
| Q8_0 | 1.85 GB |
| BF16 | 3.47 GB |
The 256K context and the 1 GB figure don’t come together. Liquid AI indicates that a full context of 128K tokens needs 1.6 GB of KV cache by itself. That’s still modest: in the same comparison, also published by the company, Qwen3.5-4B needs 4.3 GB and MiniCPM5-2B 5.6 GB for that same context. Still, budget around 1 GB for short prompts and about 3 GB if you’re really going to use 128K tokens.
The Q4_0 file is a QAD (quantization-aware) checkpoint, LFM3-1.7B-QAD-Q4_0.gguf: it’s trained to perform well at 4-bit, rather than simply rounded afterwards.
If you’re choosing a model based on your equipment’s memory, in Gemma 4 on your Mac with Ollama we did the same exercise with a larger family.
How do I run LFM3 locally with Ollama, llama.cpp, or vLLM?
All these commands come from the official model cards. Use a recent runtime: at the time of publishing this note, the GGUF model card required Ollama v0.40.12 or later and llama.cpp build b11960 or later; earlier versions don’t recognize the architecture.
How do I use LFM3 with Ollama?
Ollama downloads the GGUF directly from Hugging Face:
ollama run hf.co/LiquidAI/LFM3-1.7B-GGUF:Q4_K_M
At the time of publishing this note, Ollama was not listed among LFM3’s launch partners; this hf.co/ route is what the model card documents.
How do I use LFM3 with llama.cpp?
Install llama.cpp with brew install llama.cpp (macOS/Linux) or winget install llama.cpp (Windows), then:
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M \
--temp 0.1 --top-k 50 --repeat-penalty 1.1
To use the 4-bit QAD checkpoint:
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF \
--hf-file LFM3-1.7B-QAD-Q4_0.gguf \
--temp 0.1 --top-k 50 --repeat-penalty 1.1
How do I serve LFM3 with vLLM?
vLLM spins up an OpenAI API-compatible server:
pip install vllm
vllm serve "LiquidAI/LFM3-1.7B"
curl -X POST "http://localhost:8000/v1/chat/completions" \
-H "Content-Type: application/json" \
--data '{
"model": "LiquidAI/LFM3-1.7B",
"messages": [{"role": "user", "content": "What is the capital of France?"}]
}'
What other options are there?
- Docker Model Runner:
docker model run hf.co/LiquidAI/LFM3-1.7B - Apple Silicon: there’s an MLX version ready at
LiquidAI/LFM3-1.7B-MLX-4bit. - LM Studio: “Use this model → LM Studio” on the GGUF page (llama.cpp engine 2.65.1 or later).
- Transformers:
pip install -U transformers accelerateand loadLiquidAI/LFM3-1.7Bwithtrust_remote_code=True.
What parameters should I use with LFM3?
Liquid AI publishes two configurations:
- Chat, tool calling, agents, and code:
temperature=0.1,top_k=50,repetition_penalty=1.1 - Math and STEM:
temperature=1.0,top_p=0.95,top_k=20
There are two behaviors worth knowing about before integrating it into anything:
- Reasoning comes enabled by default. You can pass
enable_thinking=False, but the model card warns that this switch can be inconsistent, especially on tasks that require heavy reasoning. - Tool calls come out as XML (
<tool_call>,<function=...>,<parameter=...>), not JSON. If you build your own harness, parse accordingly and send tool schemas with each request. The chat template also supports deferred tools (defer_loading=True), but tool lookup you have to implement in your harness.
What is LFM3 good for, and what is it not?
It works as local AI, even without internet once the weights are downloaded, for tool calling, extraction, classification, routing, and summarization: tasks where a fast, private, cheap model beats a smart one in the cloud. Liquid AI’s main claim is that LFM3-1.7B beats its predecessor, LFM2.5-2.6B, with 35 percent fewer parameters, including a 38.42 versus 3.02 on WebArena-Verified. These are figures published by the company itself; at the time of publishing this note there was no independent evaluation.
Don’t use it as your code agent. The model card says it explicitly: not recommended for agentic programming or complex multi-step reasoning because of its size, and it suggests relying on retrieval or search for tasks that depend on knowledge. If you want a local model behind Codex CLI or Claude Code, you need one much larger: we explain it in how to use Codex CLI with a local 125B model.
Is LFM3 free for commercial use?
Yes, if your company bills less than 10 million dollars a year. The LFM Open License v1.0 is Apache 2.0 with an added condition: free commercial use ends when annual revenue exceeds that threshold, and from then on you need a commercial license from Liquid AI. Use in research and in qualified non-profit organizations has no threshold. You’re not required to publish your fine-tunes, but derivatives fall under the same license, and to redistribute you must preserve the license and attribution notices.
All LFM3 models can be fine-tuned on a single GPU; Liquid AI publishes notebooks for SFT, DPO, GRPO, and continued pretraining.
What about you? What task from your workflow would you hand off today to a 1 GB local model?
- Ollama
- llama.cpp or LM Studio
- I don’t use local models yet
- I prefer another option (tell us which)
Comment below or, if this article reached you by email, reply directly to the email: your reply gets published here.
Related: Ollaya: How to run decision models locally as an open source alternative to Jev