Instructions to use nicolasramos/odooclaw-light-1.2b-ft 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-light-1.2b-ft 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-light-1.2b-ft:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M

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-light-1.2b-ft:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M

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-light-1.2b-ft:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M

Use Docker

docker model run hf.co/nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M

LM Studio
Jan
Ollama
How to use nicolasramos/odooclaw-light-1.2b-ft with Ollama:
```
ollama run hf.co/nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M
```

Unsloth Studio

How to use nicolasramos/odooclaw-light-1.2b-ft 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-light-1.2b-ft 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-light-1.2b-ft 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-light-1.2b-ft to start chatting

How to use nicolasramos/odooclaw-light-1.2b-ft 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-light-1.2b-ft:Q4_K_M

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-light-1.2b-ft:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

OpenClaw new

How to use nicolasramos/odooclaw-light-1.2b-ft 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-light-1.2b-ft:Q4_K_M

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-light-1.2b-ft:Q4_K_M" \
  --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-light-1.2b-ft with Docker Model Runner:
```
docker model run hf.co/nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M
```

Lemonade

How to use nicolasramos/odooclaw-light-1.2b-ft with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull nicolasramos/odooclaw-light-1.2b-ft:Q4_K_M

Run and chat with the model

lemonade run user.odooclaw-light-1.2b-ft-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use nicolasramos/odooclaw-light-1.2b-ft 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-light-1.2b-ft:Q4_K_M

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-light-1.2b-ft:Q4_K_M

Run Hermes

hermes

Atomic Chat

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

OdooClaw Light

Odoo ERP assistant model with tool calling, fine-tuned from LFM2.5-1.2B-Instruct (Liquid AI). Optimized for real-world business conversation in Spanish, record creation and safe rejection of destructive operations.

Current version: v8 (2026-08-09) — canonical file, overwritten on every update.

⭐ Evaluation results (v8)

Real conversation battery — 19/20 (95%)

20 real-world business cases (greetings, typos, reading, creation, destructive):

Category	Result
Greetings / conversation	4/4 ✅
Reading (find client, count, tasks)	6/6 ✅
Creation (task, client, quotation, opportunity, vendor bill)	6/6 ✅
Destructive (delete clients, delete invoice)	2/2 ✅ (correct rejection)
Out of scope / identity	1/1 ✅

Extended business battery — 22/28 (79%)

Real typical questions: find/create partners, tasks, quotations, invoices, payments, cancellations, corrections. Known misses: "¿Cómo estás?" (calls tool), "Rectifica la factura" (should search first), "Registra el pago" (confuses with vendor_invoice), "Necesito ayuda"/"¿Qué puedes hacer?" (calls search). With the full MCP docstrings and top-5 retrieval, production performance is higher.

Creation benchmark (full schemas) — 6/6 tools, 6/6 args ✅

task, res.partner, vendor_invoice, sale_order, lead — all with complete, correct arguments.

Known limitations (documented)

Composite queries: "tareas de Juan Pérez" (user_ids), "pendientes de cerrar" (stage filter) — the 1.2B cannot generalize complex domains. Solution: synthesis tools in the MCP (odoo_find_tasks_for_user, odoo_get_task_stats, odoo_get_financial_snapshot) — the model recognizes intent, the MCP builds the domain.
Short "¿Cómo estás?" may call a tool (without the full gateway system prompt).
"Rectifica la factura X" does not chain search→edit.

🚀 Hardware performance (v8, Q4_K_M, measured 2026-08-09)

Mac Mini M1 (llama.cpp Metal)

Test	Time	Speed
Tool call (find partner)	0.66s	36.5 tok/s
Tool call (search)	0.92s	43.5 tok/s
Tool call (create task)	0.64s	36.0 tok/s
Long text (300 tok)	4.24s	60.2 tok/s

Mac Studio M1 Max (oMLX / MLX 4-bit)

Test	Time	Speed
Tool call (find partner)	0.97s	23.7 tok/s
Tool call (search)	0.31s	127.1 tok/s
Tool call (create task)	0.28s	105.0 tok/s
Long text	0.39s	139.9 tok/s

AMD Strix Halo (llama.cpp Vulkan)

Test	Time	Speed
Tool call (find partner)	0.16s	150.5 tok/s
Tool call (search)	0.22s	185.5 tok/s
Tool call (create task)	0.14s	160.6 tok/s
Long text	1.10s	231.0 tok/s

PC RTX 5070 Ti (llama.cpp CUDA)

Test	Time	Speed
Tool call (find partner)	0.35s	240.8 tok/s
Tool call (search)	0.46s	298.3 tok/s
Tool call (create task)	0.50s	300.6 tok/s
Long text	0.98s	306.1 tok/s

Intel N100 (llama.cpp CPU, 4 cores) — the client reference hardware

Config	Tool call (s)	Long text (tok/s)
4 cores / unlimited RAM	5.3-7.3	13.3
4 cores / 6 GB RAM	5.1-7.7	4.3*
4 cores / 4 GB RAM	5.3-18.9	12.5
4 cores / 2 GB RAM	6.2-6.6	11.3
3 cores / 16 GB	5.7-6.6	11.9
2 cores / 16 GB	6.2-7.0	13.5
1 core / 16 GB	5.0-7.8	12.3

*Single outlier (64s on one case); the rest of the 6GB config behaves like 16GB.

Key finding: works with 2 GB RAM and a single core — suitable for mini-PCs and low-end VPS. The bottleneck is CPU-bound: more RAM/cores barely changes latency (prompt eval dominates).

📦 Formats

GGUF: odooclaw-light-1.2b-ft-Q4_K_M.gguf (698 MB) — this repo
MLX 4-bit: repo odooclaw-light-1.2b-ft-mlx
Ollama: Modelfile included (im_start template, temp 0.0, top_k 50, repeat_penalty 1.05)

🔧 Quick deploy

# llama.cpp
llama-server -m odooclaw-light-1.2b-ft-Q4_K_M.gguf --host 0.0.0.0 --port 8085 -c 8192 --parallel 1 -t 4 --temp 0.0 --top-k 50 --repeat-penalty 1.05 --jinja

# Ollama
ollama create odooclaw-light -f Modelfile

📚 Training

Base: LiquidAI/LFM2.5-1.2B-Instruct
Dataset v11: 26,968 examples (124 odoo-mcp tools + 4 ocr-invoice + conversation + creation + composite queries + destructive)
LoRA r16, 2 epochs, batch 8 (Unsloth), ~35-50 min on RTX 5070 Ti
Full recipe: train_lfm25_12b.py

Downloads last month: 2

GGUF

Model size

1B params

Architecture

lfm2

Hardware compatibility

4-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support