Instructions to use MityaLabs/nafanya-1.7b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MityaLabs/nafanya-1.7b with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-0.8B")
model = PeftModel.from_pretrained(base_model, "MityaLabs/nafanya-1.7b")

MLX

How to use MityaLabs/nafanya-1.7b with MLX:

# Download the model from the Hub
pip install huggingface_hub[hf_xet]

hf download MityaLabs/nafanya-1.7b --local-dir nafanya-1.7b

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use MityaLabs/nafanya-1.7b 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 MityaLabs/nafanya-1.7b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf MityaLabs/nafanya-1.7b:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf MityaLabs/nafanya-1.7b:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf MityaLabs/nafanya-1.7b:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf MityaLabs/nafanya-1.7b:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf MityaLabs/nafanya-1.7b:Q4_K_M

Use Docker

docker model run hf.co/MityaLabs/nafanya-1.7b:Q4_K_M

Llama new
LM Studio
Jan
Ollama
How to use MityaLabs/nafanya-1.7b with Ollama:
```
ollama run hf.co/MityaLabs/nafanya-1.7b:Q4_K_M
```
Unsloth Desktop

How to use MityaLabs/nafanya-1.7b with Pi:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "MityaLabs/nafanya-1.7b"

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "mlx-lm": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "MityaLabs/nafanya-1.7b"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use MityaLabs/nafanya-1.7b with Docker Model Runner:
```
docker model run hf.co/MityaLabs/nafanya-1.7b:Q4_K_M
```

Lemonade

How to use MityaLabs/nafanya-1.7b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull MityaLabs/nafanya-1.7b:Q4_K_M

Run and chat with the model

lemonade run user.nafanya-1.7b-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use MityaLabs/nafanya-1.7b with Hermes Agent:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "MityaLabs/nafanya-1.7b"

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 MityaLabs/nafanya-1.7b

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use MityaLabs/nafanya-1.7b with OpenClaw:

Start the MLX server

# Install MLX LM:
uv tool install mlx-lm
# Start a local OpenAI-compatible server:
mlx_lm.server --model "MityaLabs/nafanya-1.7b"

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 "MityaLabs/nafanya-1.7b" \
  --custom-provider-id mlx-lm \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

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

nafanya-1.7b

A compact Russian tool-calling model for the offline kuzya agent on a phone (Poco F1, skladenOS). Fine-tuned from Qwen/Qwen3-1.7B with LoRA on synthetic teacher-generated data; runs as GGUF Q4_K_M through llama-server.

nafanya is the model; kuzya is the agent that runs it. The agent, its tools and its rules live in a separate repository.

What it is

Base: Qwen/Qwen3-1.7B (Apache-2.0).
Training: MLX-LM BF16 LoRA, 8 last layers, rank 16, scale 16, dropout 0.05, 100 iterations, batch 1, max sequence 1280.
Prompt during training: compact system prompt plus one-line tool descriptions (1118 tokens, 13 tools: battery, Wi-Fi status/connect, services list, service logs, service restart, sleep inhibit, volume get/set/step, mute, brightness set/step).
Data: ~610 synthetic Russian requests and refusals, generated by DeepSeek V4.1 Flash (AI-generated), filtered against the agent's frozen test sets.
Artifacts here: PEFT LoRA adapter and GGUF Q4_K_M.

Evaluation

temperature 0, compact prompt, GGUF Q4_K_M. Sets are Russian tool-selection queries with exact expected tool and arguments.

Set	Correct tool	With arguments
frozen 148-query set	131/148 (89%)	125/148 (84%)
held-out 40 (v2)	33/40 (82%)	32/40 (80%)
held-out 40 (v3)	32/40 (80%)	27/40 (68%)
held-out 40 (v4)	—	—

With the kuzya rules layer (deterministic pre-routing and argument repair in the agent, not in these weights) the same model scores 98–100% on the fresh sets: 40/40 on held-out v4 and v3 and 40/40 on v2. Do not attribute those numbers to the weights alone.

Phone (Poco F1, Q4_K_M): peak RAM 1.71 GB, load 3–4 s, warm energy 8.9 J per query with rules (30.7 J without rules), mean 2.4 s per query with rules.

Limitations

Relative vs absolute volume and brightness ("потише" with a number) is fragile.
mute_set and nosleep_set on/off can be inverted without the agent rules.
A line count can be mistaken for a service name in service_logs.
Out-of-scope requests (Wi-Fi on/off, ssh/sshd, "all services", arithmetic) may produce a tool call; the agent's allowlist and rules catch them.
Evaluated only on Russian queries and only on one phone model.

Data and licence

Weights: Apache-2.0 (same as the base model).
Training data is synthetic, generated by DeepSeek V4.1 Flash through its API under the DeepSeek Open Platform Terms of Service (release 2026-04-22, §4.2(3), which permits using outputs to train other models). It is AI-generated; no data from the phone, no personal data.
DeepSeek is mentioned only as the data source; this model is not affiliated with or endorsed by DeepSeek.

Model tree for MityaLabs/nafanya-1.7b

Base model

Qwen/Qwen3-1.7B-Base

Finetuned

Qwen/Qwen3-1.7B

Adapter

(851)

this model

MityaLabs
/

nafanya-1.7b

nafanya-1.7b

What it is

Evaluation

Limitations

Data and licence

Links

Model tree for MityaLabs/nafanya-1.7b