isox/Qwen3-14B-HomeAssistant-ru-GGUF

Qwen3-14B HomeAssistant Russian (GGUF)

Fine-tuned Qwen3-14B for Russian-language Home Assistant voice control with native tool calling.

Quick Start

# Q8_0 — best quality, needs 18GB+ VRAM (A10 24GB, RTX 3090/4090)
ollama pull hf.co/isox/Qwen3-14B-HomeAssistant-ru-GGUF:Q8_0

# Q4_K_M — good balance, needs 10GB+ VRAM (RTX 3060 12GB+)
ollama pull hf.co/isox/Qwen3-14B-HomeAssistant-ru-GGUF:Q4_K_M

Custom Modelfile (recommended)

The included Modelfile sets recommended parameters (32K context, tool calling template, optimized sampling):

# Download GGUF + Modelfile, then create with custom settings
ollama create qwen3-14b-ha-ru -f Modelfile

Recommended Ollama environment for NVIDIA A10 24GB:

OLLAMA_FLASH_ATTENTION=1
OLLAMA_KV_CACHE_TYPE=q8_0
OLLAMA_NUM_PARALLEL=1

Model Details

Parameter	Value
Base model	Qwen/Qwen3-14B
Method	QLoRA → LoRA merge → GGUF
Training data	214,339 examples (Russian smart home commands)
LoRA rank	64 (alpha=128)
Training	1 epoch, 3,350 steps, 43h on NVIDIA H100 SXM 80GB
Final loss	0.2178
LoRA adapter	isox/Qwen3-14B-LoRA-HomeAssistant-ru

Available Quantizations

File	Quant	Size	VRAM	Speed (A10)
Qwen3-14B-HomeAssistant-ru-Q8_0.gguf	Q8_0	15 GB	~18 GB	~29 tok/s
Qwen3-14B-HomeAssistant-ru-Q4_K_M.gguf	Q4_K_M	8.4 GB	~10 GB	~35 tok/s

Supported Tools (Home Assistant)

The model was trained on the following Home Assistant service tool calls:

• HassTurnOn / HassTurnOff — turn devices on/off
• HassSetTemperature — set thermostat/AC temperature
• HassStartTimer — start timers
• HassLightSet — set brightness/color
• HassOpenCover / HassCloseCover — blinds/curtains
• HassMediaPause / HassMediaNext / HassVolumeSet — media control
• HassLockLock / HassLockUnlock — door locks
• HassVacuumStart / HassVacuumReturnToBase — robot vacuum
• HassGetState — query device state

Usage with Home Assistant

This model is designed for use with Home LLM integration. It responds to Russian voice commands and produces Qwen3 tool calls in the correct format.

Example

User: Включи свет в гостиной
Model: HassTurnOn(name="Свет в гостиной")

User: Поставь кондиционер на 22 градуса
Model: HassSetTemperature(name="Кондиционер", temperature=22)

User: Привет, как дела?
Model: Привет! Всё отлично, чем могу помочь?

Training Data

214,339 examples generated from Russian device/action piles covering:

• 13 assistant personas (formal, friendly, sarcastic, slang, etc.)
• 2,016 device names
• 1,569 response templates
• 1,330 specific action phrases
• 646 templated action patterns

Dataset: isox/home-assistant-russian-train (private)