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 AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("LiquidAI/LFM3-1.7B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("LiquidAI/LFM3-1.7B", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

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
```

Five 1.7B role specialists fine-tuned from LFM3-1.7B-Base

#28

by SkyIsNotGreen - opened 7 days ago

Discussion

SkyIsNotGreen

7 days ago

•

edited 7 days ago

Hello,

I've been working on Hive, a local harness for role-specialized models. I fine-tuned five 1.7B models from LiquidAI/LFM3-1.7B-Base. Each one covers a different seat in the harness instead of trying to be a general chat model.

Collection: Hive series
Harness: sky-is-green/hive

The five models are:

Hive Conductor 1.7B: controller and planner. 59/60 on controller, 69/77 planner on HF, 66/77 on GGUF.
Hive Reader 1.7B: mechanics and RAG. 1.000 mechanics and 1.000 RAG on HF, 0.975 and 1.000 on GGUF.
Hive Auditor 1.7B: placement and critic. 0.957 and 0.947 on HF, 0.935 and 0.921 on GGUF.
Hive Dispatcher 1.7B: dispatcher. 0.950 on HF, 0.925 on GGUF.
Hive Tool Manager 1.7B: agency. 0.977 on HF, 0.930 on GGUF.

The scores come from my held-out role suites with greedy decoding. The GGUF rows use Q4_K_M. Each task family has 38 to 77 examples; I put the task files, checkers and eval records in the repo.

Hive runs a loop: solve, detect, repair, escalate. The seats fail in different ways, so I trained one small model for each one. A 1.7B model is cheap enough that I can keep several loaded on one workstation, even alongside a 27b-class generative model, and the checkers give me a signal that doesn't rely on it being there.

The five seats do different jobs. The Conductor assigns work, respects capacity and dependency constraints, and produces an ordered plan with risks. The Reader extracts invoice or contract fields, or answers only from a supplied passage. The Auditor fits a model into vRAM/KV budgets, or validates a proposed assignment, schedule or order. The Dispatcher assigns jobs to workers and returns a dependency, or deadline-aware schedule. The Tool Manager chooses the right tool and orders multi-tool calls before an agent runs them.

I used synthetic data that passed deterministic checkers, with no user data, and trained them on one 20 GB AMD card. I released merged Transformers checkpoints plus Q4_K_M GGUF files, and added a use-cases section with concrete examples to each model card.

I'm sharing this because I think narrow role fine-tunes are a useful pattern for local models. I'd like feedback on the role boundaries and the eval design. I'm also happy to share the data-generation, LoRA and GGUF recipes.

I should be clear about the limits. These are independent community fine-tunes, not LiquidAI releases. They are role specialists, not chat models. The numbers are from my own held-out suites, not a general claim about model quality. For open conversation, use the LFM3 base or instruct model.

I'm releasing them as derivatives of LiquidAI/LFM3-1.7B-Base under the LFM Open License v1.0. Commercial use by an entity above USD 10M annual revenue is not licensed under LFM 1.0. The base model is by Liquid AI.

Thanks for releasing it!

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