Liquid Foundation Models

https://docs.liquid.ai
Admin
Liquid Foundation Models are hybrid AI models developed by Liquid AI for efficient deployment across...
Tokens:211,825
Snippets:1,689
Trust Score:9.8
Update:1 day ago
Show doc for...
Tokens:
Raw
### Install TypeSafe SDK

Source: https://docs.liquid.ai/guides/decision-model-guide

Install the TypeSafe SDK for decision model examples.

```bash
pip install typesafe-sdk
```

--------------------------------

### One-time setup for liquid-finetune

Source: https://docs.liquid.ai/examples/customize-models/home-assistant

Clones and installs liquid-finetune, then authenticates with Hugging Face and Modal.

```bash
git clone https://github.com/Liquid4All/liquid-finetune
cd liquid-finetune && uv sync && cd -
huggingface-cli login
modal setup
```

--------------------------------

### Development setup: install huggingface-hub and push model to Android

Source: https://docs.liquid.ai/deployment/on-device/llama-cpp/mobile

Commands to install the Hugging Face hub CLI, download the model, and push it to an Android device for development.

```bash
uv pip install huggingface-hub
hf download LiquidAI/LFM2.5-1.2B-Instruct-GGUF LFM2.5-1.2B-Instruct-Q4_0.gguf --local-dir .
adb push LFM2.5-1.2B-Instruct-Q4_0.gguf /data/local/tmp/   # Android
```

--------------------------------

### Install PyTorch, Transformers, and image helper libraries

Source: https://docs.liquid.ai/lfm/key-concepts/vision-capabilities

Install the required packages for LFM2.5-VL examples: PyTorch, torchvision, accelerate, Pillow, opencv-python, and Transformers (>=5.10.1).

```bash
# Install PyTorch and acceleration libraries
%pip install -q torch torchvision accelerate

# Install image helper libraries
%pip install -q pillow opencv-python

# Install Transformers
%pip install -q "transformers>=5.10.1"
```

--------------------------------

### Install Hermes Agent

Source: https://docs.liquid.ai/examples/agent-harnesses

Install Hermes Agent using the official install script and run the setup wizard.

```bash
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
```

--------------------------------

### Install make

Source: https://docs.liquid.ai/examples/customize-models/car-maker-identification

Install make using apt-get on Linux.

```bash
sudo apt-get install make
```

--------------------------------

### Run locally with npm install and npm run dev

Source: https://docs.liquid.ai/examples/web/hand-voice-racer

Install dependencies and start the development server. Requires Node.js 18+ and a browser with WebGPU support (Chrome/Edge 113+).

```bash
npm install
npm run dev
```

--------------------------------

### Install liquid-finetune CLI as uv tool

Source: https://docs.liquid.ai/lfm/fine-tuning/liquid-finetune

Install the CLI for repeated use from any directory. After installation, run `liquid-finetune` with an absolute path to a config file.

```bash
uv tool install git+https://github.com/Liquid4All/liquid-finetune.git
liquid-finetune /absolute/path/to/config.yaml
```

--------------------------------

### Manually download model and mmproj files

Source: https://docs.liquid.ai/deployment/on-device/llama-cpp

Installs huggingface-hub and downloads the model and mmproj files for manual setup.

```bash
uv pip install huggingface-hub
hf download LiquidAI/LFM2-VL-1.6B-GGUF LFM2-VL-1.6B-Q8_0.gguf --local-dir .
hf download LiquidAI/LFM2-VL-1.6B-GGUF mmproj-LFM2-VL-1.6B-Q8_0.gguf --local-dir .
```

--------------------------------

### Start OpenAI-compatible server and use OpenAI Python client

Source: https://docs.liquid.ai/deployment/on-device/lm-studio

Start the server from the Developer tab, then use the OpenAI Python client with base_url http://localhost:1234/v1 and api_key 'not-needed'. The model name can be any string. The example sends a chat completion with temperature 0.1, max_tokens 512, and extra_body parameters top_k and repeat_penalty.

```python
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:1234/v1",
    api_key="not-needed"
)

response = client.chat.completions.create(
    model="local-model",  # Any string works
    messages=[
        {"role": "user", "content": "What is machine learning?"}
    ],
    temperature=0.1,
    max_tokens=512,
    extra_body={"top_k": 50, "repeat_penalty": 1.05},
)
print(response.choices[0].message.content)
```

--------------------------------

### Install OpenClaw

Source: https://docs.liquid.ai/examples/agent-harnesses

Install OpenClaw using the official install script and onboard with the daemon.

```bash
curl -fsSL https://openclaw.ai/install.sh | bash   # macOS / Linux
openclaw onboard --install-daemon
```

--------------------------------

### Start llama-server with LFM2.5-1.2B-Instruct-GGUF

Source: https://docs.liquid.ai/deployment/on-device/llama-cpp/chat

Starts llama-server with the specified model, context length, port, and chat template. The -hf flag downloads the GGUF from Hugging Face on first run; :Q4_K_M selects the quantization. Use -m path/to/model.gguf for a local file. -c sets the context length; larger contexts cost memory linearly. --jinja uses the chat template embedded in the GGUF, required for tool calling. -ngl 99 offloads all layers to the GPU when available. -np 4 serves up to four concurrent requests, each getting -c / 4 tokens of context.

```bash
llama-server -hf LiquidAI/LFM2.5-1.2B-Instruct-GGUF:Q4_K_M -c 4096 --port 8080 --jinja
```

--------------------------------

### Install uv on Windows

Source: https://docs.liquid.ai/examples/customize-models/car-maker-identification

Install uv using PowerShell on Windows.

```powershell
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
```

--------------------------------

### Install PyLate

Source: https://docs.liquid.ai/lfm/models/lfm2-colbert-350m

Install the PyLate library for indexing, retrieval, and reranking.

```bash
pip install -U pylate
```

--------------------------------

### Start the development server

Source: https://docs.liquid.ai/examples/web/vl-webgpu-demo

Start the Vite development server. The application will be accessible at http://localhost:5173.

```bash
npm run dev
```

--------------------------------

### Launch the demo

Source: https://docs.liquid.ai/examples/laptop-examples/audio-car-cockpit

Starts the audio server and the main serve process in parallel using make -j2.

```bash
make -j2 audioserver serve
```

--------------------------------

### Prefill for structured output example

Source: https://docs.liquid.ai/lfm/key-concepts/text-generation-and-prompting

Starts the model with a specific format or structure (e.g., JSON opening brace) to guide it toward structured outputs. For structured generation with schema validation, see Outlines.

```python
messages = [
    {"role": "system", "content": "Extract information and return as JSON."},
    {"role": "user", "content": "Extract the name and age from: John is 30 years old."},
    {"role": "assistant", "content": "{\n  \"name\": "}  # Prefill with JSON structure
]
```

--------------------------------

### Setup Python environment

Source: https://docs.liquid.ai/examples/laptop-examples/audio-car-cockpit

Clones the cookbook repository and runs make setup to prepare the environment for the audio car cockpit demo.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/audio-car-cockpit
make setup
```

--------------------------------

### Clone repository and navigate to audio-webgpu-demo

Source: https://docs.liquid.ai/examples/web/audio-webgpu-demo

Clone the cookbook repository and change into the audio-webgpu-demo directory. This is the first step in the Quick Start workflow.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/audio-webgpu-demo
```

--------------------------------

### Install vLLM

Source: https://docs.liquid.ai/examples/agent-harnesses

Install vLLM for GPU servers. See the vLLM guide for more details.

```bash
pip install vllm
```

--------------------------------

### Clone the repository and navigate to the example directory

Source: https://docs.liquid.ai/examples/laptop-examples/audio-to-text-in-real-time

Clones the cookbook repository and changes into the audio-transcription-cli example directory.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/audio-transcription-cli
```

--------------------------------

### Clone and build the project

Source: https://docs.liquid.ai/examples/laptop-examples/flight-search-assistant

Clone the cookbook repository, navigate to the flight-search-assistant example directory, and sync dependencies with uv.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/flight-search-assistant
uv sync
```

--------------------------------

### Bash code block example

Source: https://docs.liquid.ai/lfm/help/contributing

Example of a bash code block for installing dependencies with pip.

```markdown
```bash
pip install transformers torch
```
```

--------------------------------

### Install SGLang

Source: https://docs.liquid.ai/examples/agent-harnesses

Install SGLang version 0.5.10 or later for GPU servers. See the SGLang guide for more details.

```bash
uv pip install "sglang>=0.5.10"
```

--------------------------------

### Clone the repository and navigate to the demo directory

Source: https://docs.liquid.ai/examples/web/vl-webgpu-demo

Clone the cookbook repository and change into the vl-webgpu-demo example directory.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/vl-webgpu-demo
```

--------------------------------

### Install numpy and requests for GGUF embedding client

Source: https://docs.liquid.ai/lfm/models/lfm25-embedding-350m

Install the Python packages needed for the GGUF embedding client example.

```bash
pip install numpy requests
```

--------------------------------

### Clone the repository and navigate to the example directory

Source: https://docs.liquid.ai/examples/customize-models/car-maker-identification

Clones the cookbook repository and changes into the car-maker-identification example directory.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/car-maker-identification
```

--------------------------------

### Install OpenAI SDK and pydantic

Source: https://docs.liquid.ai/guides/decision-model-guide

Install the OpenAI SDK and pydantic for LLM examples. The OpenAI SDK works with any OpenAI-compatible endpoint.

```bash
pip install openai pydantic
```

--------------------------------

### Few-shot prompting example

Source: https://docs.liquid.ai/lfm/key-concepts/text-generation-and-prompting

Provides example interactions to guide the model's behavior. The model learns from the conversation history and applies patterns to new inputs.

```python
messages = [
    {"role": "system", "content": "You are a helpful assistant that formats dates."},
    {"role": "user", "content": "2024-01-15"},
    {"role": "assistant", "content": "January 15, 2024"},
    {"role": "user", "content": "2024-12-25"},
    {"role": "assistant", "content": "December 25, 2024"},
    {"role": "user", "content": "2024-03-08"}  # Model follows the pattern
]
```

--------------------------------

### Launch fine-tuning with liquid-finetune

Source: https://docs.liquid.ai/examples/customize-models/wildfire-prevention

Starts the fine-tuning job using the prepared configuration file. Run from the liquid-finetune directory.

```bash
cd liquid-finetune && uv run liquid-finetune ../configs/wildfire_finetune_modal.yaml
```

--------------------------------

### Copy .env.example to .env

Source: https://docs.liquid.ai/examples/laptop-examples/build-a-game-with-decision-models

Create a .env file from the template.

```bash
cp .env.example .env
```

--------------------------------

### Install llama.cpp via Homebrew or winget

Source: https://docs.liquid.ai/examples/agent-harnesses

Install llama.cpp on macOS using Homebrew or on Windows using winget. For Linux and build-from-source options, see the llama.cpp guide.

```bash
brew install llama.cpp        # macOS
winget install llama.cpp      # Windows
```

--------------------------------

### Clone the repository and navigate to the example directory

Source: https://docs.liquid.ai/examples/laptop-examples/lfm2-english-to-korean

Clones the Liquid AI cookbook repository and changes into the lfm2-english-to-korean example directory.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/lfm2-english-to-korean
```

--------------------------------

### Kick off training on Modal

Source: https://docs.liquid.ai/examples/customize-models/home-assistant

Runs 5 epochs of LoRA SFT on an H100. Takes a few minutes and costs roughly $1.50.

```bash
cd liquid-finetune
uv run liquid-finetune ../finetune/configs/LFM2-350M.yaml
```

--------------------------------

### Start llama-server with local model and mmproj files

Source: https://docs.liquid.ai/deployment/on-device/llama-cpp/multimodal

Starts the server using locally downloaded model and projector files. First downloads the files with hf download, then runs llama-server with the --mmproj flag. Useful flags include --image-max-tokens and --no-mmproj-offload.

```bash
hf download LiquidAI/LFM2.5-VL-1.6B-GGUF LFM2.5-VL-1.6B-Q4_0.gguf mmproj-LFM2.5-VL-1.6b-Q8_0.gguf --local-dir .

llama-server -m LFM2.5-VL-1.6B-Q4_0.gguf --mmproj mmproj-LFM2.5-VL-1.6b-Q8_0.gguf \
  -c 4096 --port 8080 --jinja -ngl 99
```

--------------------------------

### Load LFM2.5-1.2B-Instruct with transformers

Source: https://docs.liquid.ai/lfm/models/lfm2-1.2b-tool

Load the LFM2.5-1.2B-Instruct model using the transformers library. The model is loaded with device_map='auto'. See the Tool Use guide for complete examples.

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "LiquidAI/LFM2.5-1.2B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

# See the Tool Use guide for complete examples
```

--------------------------------

### Get per-token embeddings via curl

Source: https://docs.liquid.ai/lfm/models/lfm25-encoder-350m

Sends a request to the running llama-server to obtain embeddings for the input text. Run in a separate shell after starting the server.

```bash
curl -s http://localhost:8080/embedding -d '{"content": "hello world"}'
```

--------------------------------

### Start Claude Code or Codex from repo root

Source: https://docs.liquid.ai/lfm/fine-tuning/liquid-finetune

Launches Claude Code or Codex from the repository root. The repo includes AGENTS.md, CLAUDE.md, and skills to guide the agent.

```bash
claude
# or
codex
```

--------------------------------

### Run Pi

Source: https://docs.liquid.ai/examples/agent-harnesses

Start Pi and then select the model with /model.

```bash
pi
```

--------------------------------

### Run Gradio demo

Source: https://docs.liquid.ai/lfm/models/lfm2-audio-1.5b

Starts the Gradio demo webserver on http://localhost:7860/.

```bash
liquid-audio-demo
# Starts webserver on http://localhost:7860/
```

--------------------------------

### Run invoice parser with make commands

Source: https://docs.liquid.ai/examples/laptop-examples/invoice-extractor-tool-with-liquid-nanos

If make is installed, these commands run the application: 'make run' starts watch mode and 'make process' runs one-shot process mode.

```bash
make run       # watch mode
make process   # one-shot process mode
```

--------------------------------

### Clone repository and sync dependencies

Source: https://docs.liquid.ai/examples/customize-models/car-maker-identification

Clone the cookbook repository, navigate to the car-maker-identification example, and sync dependencies with uv.

```bash
git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/car-maker-identification
uv sync
```
Tokens:
Raw