Instructions to use LiquidAI/LFM3-1.7B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LiquidAI/LFM3-1.7B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="LiquidAI/LFM3-1.7B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM3-1.7B", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use LiquidAI/LFM3-1.7B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LiquidAI/LFM3-1.7B"
# Call the server using curl (OpenAI-compatible API):
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?"
			}
		]
	}'

Use Docker

docker model run hf.co/LiquidAI/LFM3-1.7B

SGLang

How to use LiquidAI/LFM3-1.7B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "LiquidAI/LFM3-1.7B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LiquidAI/LFM3-1.7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "LiquidAI/LFM3-1.7B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LiquidAI/LFM3-1.7B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use LiquidAI/LFM3-1.7B with Docker Model Runner:
```
docker model run hf.co/LiquidAI/LFM3-1.7B
```

Unexpected </think> with enable_thinking=False — expected behavior?

#20

by DTwingy - opened 1 day ago

Discussion

DTwingy

1 day ago

•

edited 1 day ago

Hello! I’m seeing a closing tag inside generated text despite disabling thinking.

Model: LiquidAI/LFM3-1.7B

Messages:
System: Answer factual questions briefly.
User: What is a Renault Twingo? Answer in one sentence.

Settings:
enable_thinking=False
temperature=0.1
top_k=50
top_p=1.0
repetition_penalty=1.1

Observed output includes:

…hatchback or city car.</think>A Renault Twingo is…

This reproduced in fresh llama.cpp sessions using the official Q4_K_M GGUF. Our runtime includes the fix from llama.cpp PR #31374, whose regression tests passed.

It also appeared using the original BF16 checkpoint with Transformers 5.22.1 and PyTorch 2.8.0 CPU.

Reference-test limitations:

Linear calculations were promoted to FP32, then rounded back to BF16 activations for CPU feasibility; weights were unchanged.
The reference run stopped at its 300-second limit after 64 generated tokens. The tag appeared before that stop.
This is one prompt and seed, not a comprehensive evaluation.

Is this expected raw-output behavior requiring a specific response parser, or an issue with thinking-disabled generation? What is the recommended handling for streaming text into speech?

I can provide pinned revisions, token IDs and raw response logs if helpful.

mlabonne

Liquid AI org about 19 hours ago

Hey, yes, leakage of thinking traces is expected because it was trained as a reasoning model. You can experiment with different generation parameters to help prevent this, but unfortunately it's not 100% reliable. We'll work on this for the next release.

DTwingy

about 18 hours ago

Thanks for the explanation! We’re experimenting with LFM3-1.7B for a local voice assistant on a Raspberry Pi 5. Are there generation settings you’d recommend in the meantime? Looking forward to the next release and happy to help test it!

mlabonne

Liquid AI org about 8 hours ago

I don't have specific recommendations because it depends on your use case, so you may need to experiment with a few configurations to see what works best for you. Thanks!

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment