Instructions to use LUNAV/mai-nano-router with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LUNAV/mai-nano-router with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="LUNAV/mai-nano-router",
	filename="mai-nano-router-v4sft-f16.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use LUNAV/mai-nano-router 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 LUNAV/mai-nano-router:F16
# Run inference directly in the terminal:
llama cli -hf LUNAV/mai-nano-router:F16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf LUNAV/mai-nano-router:F16
# Run inference directly in the terminal:
llama cli -hf LUNAV/mai-nano-router: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 LUNAV/mai-nano-router:F16
# Run inference directly in the terminal:
./llama-cli -hf LUNAV/mai-nano-router: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 LUNAV/mai-nano-router:F16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf LUNAV/mai-nano-router:F16

Use Docker

docker model run hf.co/LUNAV/mai-nano-router:F16

LM Studio
Jan

vLLM

How to use LUNAV/mai-nano-router with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/LUNAV/mai-nano-router:F16

Ollama
How to use LUNAV/mai-nano-router with Ollama:
```
ollama run hf.co/LUNAV/mai-nano-router:F16
```

Unsloth Studio

How to use LUNAV/mai-nano-router 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 LUNAV/mai-nano-router 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 LUNAV/mai-nano-router to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for LUNAV/mai-nano-router to start chatting

How to use LUNAV/mai-nano-router with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf LUNAV/mai-nano-router: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": "LUNAV/mai-nano-router:F16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use LUNAV/mai-nano-router with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf LUNAV/mai-nano-router: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 LUNAV/mai-nano-router:F16

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use LUNAV/mai-nano-router with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf LUNAV/mai-nano-router: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 "LUNAV/mai-nano-router: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 LUNAV/mai-nano-router with Docker Model Runner:
```
docker model run hf.co/LUNAV/mai-nano-router:F16
```

Lemonade

How to use LUNAV/mai-nano-router with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull LUNAV/mai-nano-router:F16

Run and chat with the model

lemonade run user.mai-nano-router-F16

List all available models

lemonade list

MAI-nano Router (v4 SFT · GGUF)

On-device intent router + slot extractor for myMAI, a macOS natural-language launcher. Given a single user query it emits one line of JSON — {"intent": ..., "confidence": ..., "slots": {...}} — that the app turns into a native action (Finder search, Calendar query, iMessage/Mail, Music, Web search, or system.unknown).

Base model: Qwen3.5-0.8B-Base (774M params, qwen3_5 linear-attention hybrid)
Method: LoRA SFT (r=32, α=64, dropout 0.05, all-linear targets, 2 epochs, MPS/Apple Silicon), then merged
Task: intent classification + slot extraction only — a downstream native Validator re-checks every field
Languages: English, Korean, Japanese, Chinese
Runtime: llama.cpp llama-server with grammar-forced JSON output

Files

File	Quant	Size	Notes
`mai-nano-router-v4sft-q8_0.gguf`	Q8_0	811 MB	Recommended — the exact file myMAI ships on-device
`mai-nano-router-v4sft-f16.gguf`	F16	1.5 GB	Full-precision, for re-quantizing to other formats

Intents & slots

Six intents: finder.search, calendar.query, message.send, music.play, web.search, system.unknown. Slots are intent-specific (e.g. keyword, file_type, location, date_range for search; recipient, body, channel for messages). The full slot contract lives in the JSON schema below.

Evaluation (v4 SFT, 120-example multilingual holdout)

Metric	Value
Intent accuracy	99.2%
JSON validity	100%
Slot F1	0.79 (P 0.80 / R 0.78)
Hallucinated-slot rate	4.2%
Confidence MAE	0.049

Per-language intent accuracy: en 100% · ja 100% · ko 100% · zh 96.7%.

Usage

The model is trained on a canonical system prompt and expects JSON-schema-constrained decoding. Start a server:

llama-server -m mai-nano-router-v4sft-q8_0.gguf --port 8765 -ngl 99 --ctx-size 2048

System prompt (byte-identical to what the model was trained on — send it verbatim):

You are the intent router for myMAI, a macOS natural-language launcher. Read the user's single query and output ONLY one single-line JSON object: {"intent":"<id>","confidence":<0..1>,"slots":{...}}. Intent ids: finder.search, calendar.query, message.send, music.play, web.search, system.unknown. Never add keys, text, or explanations outside the JSON.

Request (OpenAI-compatible endpoint) with the enforced schema:

curl http://127.0.0.1:8765/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "messages": [
    {"role": "system", "content": "<canonical system prompt above>"},
    {"role": "user", "content": "한글파일 찾아줘"}
  ],
  "response_format": {
    "type": "json_schema",
    "json_schema": {"name": "route", "schema": {
      "type": "object",
      "properties": {
        "intent": {"type": "string", "enum": ["finder.search","calendar.query","message.send","music.play","web.search","system.unknown"]},
        "confidence": {"type": "number", "minimum": 0, "maximum": 1},
        "slots": {"type": "object", "properties": {
          "keyword": {"type": "string"}, "file_type": {"type": "string"}, "location": {"type": "string"},
          "date_range": {"type": "string"}, "recipient": {"type": "string"}, "body": {"type": "string"},
          "channel": {"type": "string", "enum": ["imessage","sms","mail"]},
          "action": {"type": "string", "enum": ["play","pause","next","prev"]},
          "query": {"type": "string"}, "want_answer": {"type": "boolean"}
        }, "additionalProperties": false}
      },
      "required": ["intent","confidence","slots"], "additionalProperties": false
    }}
  }
}'

Example output:

{"intent":"finder.search","confidence":0.92,"slots":{"file_type":"hwp"}}

Intended use & limitations

Scope: short launcher-style queries in en/ko/ja/zh. It classifies intent and extracts slots only — it does not execute actions, and myMAI re-validates every slot natively (enum grounding, hallucination checks) before acting.
Out of scope: open-ended chat, multi-step requests, or capabilities outside the six intents → these are expected to route to system.unknown.
Always decode with the JSON schema above; free-form decoding is unsupported.

License

Apache-2.0, inherited from the Qwen3.5-0.8B-Base base model.

Downloads last month: 32

GGUF

Model size

0.8B params

Architecture

qwen35

Hardware compatibility

8-bit

16-bit

Model tree for LUNAV/mai-nano-router

Base model

Qwen/Qwen3.5-0.8B-Base

Quantized

(31)

this model