Instructions to use nicolasramos/odooclaw-vision-450m with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use nicolasramos/odooclaw-vision-450m with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf nicolasramos/odooclaw-vision-450m:F16
# Run inference directly in the terminal:
llama cli -hf nicolasramos/odooclaw-vision-450m:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf nicolasramos/odooclaw-vision-450m:F16
# Run inference directly in the terminal:
llama cli -hf nicolasramos/odooclaw-vision-450m:F16

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf nicolasramos/odooclaw-vision-450m:F16
# Run inference directly in the terminal:
./llama-cli -hf nicolasramos/odooclaw-vision-450m:F16

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf nicolasramos/odooclaw-vision-450m:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf nicolasramos/odooclaw-vision-450m:F16

Use Docker

docker model run hf.co/nicolasramos/odooclaw-vision-450m:F16

LM Studio
Jan

vLLM

How to use nicolasramos/odooclaw-vision-450m with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nicolasramos/odooclaw-vision-450m"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nicolasramos/odooclaw-vision-450m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nicolasramos/odooclaw-vision-450m:F16

Ollama
How to use nicolasramos/odooclaw-vision-450m with Ollama:
```
ollama run hf.co/nicolasramos/odooclaw-vision-450m:F16
```

Unsloth Studio

How to use nicolasramos/odooclaw-vision-450m with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for nicolasramos/odooclaw-vision-450m to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for nicolasramos/odooclaw-vision-450m to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for nicolasramos/odooclaw-vision-450m to start chatting

How to use nicolasramos/odooclaw-vision-450m with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf nicolasramos/odooclaw-vision-450m:F16

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "nicolasramos/odooclaw-vision-450m:F16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

OpenClaw new

How to use nicolasramos/odooclaw-vision-450m with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf nicolasramos/odooclaw-vision-450m:F16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "nicolasramos/odooclaw-vision-450m:F16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use nicolasramos/odooclaw-vision-450m with Docker Model Runner:
```
docker model run hf.co/nicolasramos/odooclaw-vision-450m:F16
```

Lemonade

How to use nicolasramos/odooclaw-vision-450m with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull nicolasramos/odooclaw-vision-450m:F16

Run and chat with the model

lemonade run user.odooclaw-vision-450m-F16

List all available models

lemonade list

Hermes Agent

How to use nicolasramos/odooclaw-vision-450m with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf nicolasramos/odooclaw-vision-450m:F16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default nicolasramos/odooclaw-vision-450m:F16

Run Hermes

hermes

Atomic Chat

OdooClaw Vision 450M

Local invoice extraction for Odoo — read PDFs, create vendor bills, zero cloud.

Fine-tuned LFM2.5-VL-450M-Extract (Liquid AI) for document extraction inside Odoo. This is the vision companion to OdooClaw Light 1.2B FT: the chat model handles the conversation, this model reads the invoices.

What it does

Extracts structured data from invoice images: vendor, VAT, invoice number, date, client, subtotal, tax, total
Feeds the OdooClaw ocr-invoice MCP: PDF → image → this model → vendor bill created in Odoo
100% local — the default in production (replaces gpt-4o-mini cloud)
Verified end-to-end: PDF attachment → vendor bill created in Odoo (account.move), partner auto-matched, attachment linked

Why this model

450M params — the smallest model of its class, runs anywhere (even the Mac Mini M1)
Specialized in document extraction (Liquid's 2,000-invoice benchmark):
- JSON Validity: 98.9% | F1: 98.8% | VLM Judge: 84.5
- Beats models 4x its size (InternVL3.5-2B: 99.6/99.2/87.7)
OCRBench: 684
128K context

Files — WHICH QUANTIZATION TO USE

File	Size	Verdict
`odooclaw-vision-q6km.gguf`	280MB	✅ RECOMMENDED — verified 8/8 (100%) extraction success on CPU
`odooclaw-vision-q4km.gguf`	219MB	⚠️ Q4_K_M causes repetition loops on CPU — JSON never closes (0/8)
`mmproj-odooclaw-vision-f16.gguf`	189MB	Required (vision tower) — use with either

Important empirical finding: at 450M scale, Q4_K_M quantization is too aggressive for stable CPU extraction — the model enters repetition loops ("imquestos imquestos ...") and never emits valid JSON. Q6_K is the sweet spot: 100% success in 8/8 trials, only 61MB larger. Use Q6_K.

Verified extraction (real tests, 2026-08-07, N100 CPU, llama.cpp build 525)

Test 1 — generated ES invoice (DejaVu font, pdftoppm 170 DPI):

{"vendor_name": "Suministros Industriales García S.L.", "invoice_number": "b76543210",
 "date": "05/07/2026", "total": "2541.00 euro", "tax": "441.00 euro"}

Test 2 — Odoo demo vendor bill → CREATED IN ODOO:

{
  "success": true, "move_id": 26, "partner_id": 45, "attachment_linked": true,
  "invoice_data": {
    "partner_name": "Azure Interior Solutions Private Limited",
    "invoice_date": "2018-09-30", "ref": "692130859",
    "amount_tax": 541.1, "amount_total": 541.1
  }
}

Key fields (vendor, date, ref, totals) extracted and a real vendor bill was created in Odoo with the PDF attached — fully local.

Usage (llama.cpp)

# IMPORTANT: llama.cpp build 525+ required (LFM2-VL support)
llama-server -m odooclaw-vision-q6km.gguf \
  --mmproj mmproj-odooclaw-vision-f16.gguf \
  --port 8093 -c 8192 -t 3 \
  --temp 0.0 --top-k 50 --repeat-penalty 1.05 --jinja

Send the invoice image as a data URL (data:image/png;base64,...) with the image BEFORE the text prompt — verified critical for the 450M (text-first → garbage output).

Prompt notes (from testing)

Keep the prompt short: "Extract invoice data as JSON only. Return JSON with vendor_name, invoice_number, date, total, tax."
No inline JSON schema: showing the schema makes the 450M echo it back empty
max_tokens ~150: enough for the JSON, prevents repetition loops
Image first, text last in the content array

Conversion notes

Built from the official Extract PyTorch weights: language model exported separately (709MB f16 → quantize to Q6_K) + vision tower as mmproj (189MB). MLX conversion not yet available (mlx_lm 0.29 drops the vision tower).