Instructions to use NotRIK/Veya-4B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use NotRIK/Veya-4B-GGUF with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="NotRIK/Veya-4B-GGUF")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("NotRIK/Veya-4B-GGUF", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use NotRIK/Veya-4B-GGUF 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 NotRIK/Veya-4B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf NotRIK/Veya-4B-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf NotRIK/Veya-4B-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf NotRIK/Veya-4B-GGUF: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 NotRIK/Veya-4B-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf NotRIK/Veya-4B-GGUF:Q4_K_M

Use Docker

docker model run hf.co/NotRIK/Veya-4B-GGUF:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use NotRIK/Veya-4B-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "NotRIK/Veya-4B-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "NotRIK/Veya-4B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/NotRIK/Veya-4B-GGUF:Q4_K_M

SGLang

How to use NotRIK/Veya-4B-GGUF with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "NotRIK/Veya-4B-GGUF" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "NotRIK/Veya-4B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "NotRIK/Veya-4B-GGUF" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "NotRIK/Veya-4B-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use NotRIK/Veya-4B-GGUF with Ollama:
```
ollama run hf.co/NotRIK/Veya-4B-GGUF:Q4_K_M
```
Unsloth Desktop
Docker Model Runner
How to use NotRIK/Veya-4B-GGUF with Docker Model Runner:
```
docker model run hf.co/NotRIK/Veya-4B-GGUF:Q4_K_M
```

Lemonade

How to use NotRIK/Veya-4B-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull NotRIK/Veya-4B-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Veya-4B-GGUF-Q4_K_M

List all available models

lemonade list

Atomic Chat

Veya 4B

Created by NotRIK

Veya 4B is a fine-tuned language model based on Qwen3-4B-Base.

It is not trained from scratch. It is designed as the language-model component of the Veya local AI assistant project.

Model details

Property	Details
Model	Veya 4B
Creator	NotRIK
Base model	Qwen3-4B-Base
Architecture	Qwen3 dense
Parameters	Approximately 4 billion
Fine-tuning	QLoRA
Training GPU	Google Colab NVIDIA T4
Base quantization during training	4-bit NF4
Compute precision	FP16 on T4
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Maximum sequence length	1,024
Configured epochs	3
Learning rate	2e-4
Random seed	42

Training

The model was fine-tuned using:

PyTorch
Hugging Face Transformers
PEFT
bitsandbytes
QLoRA with NF4
Gradient checkpointing
Paged 8-bit AdamW optimizer
Early stopping and checkpoint restoration

The LoRA adapter was merged into the FP16 base model.

The training notebook includes dataset validation and a held-out 28-item automatic evaluation benchmark.

The bundled dataset is a small seed dataset intended for pipeline validation. No claim of large-scale training or benchmark-proven improvement is made here.

Thinking mode

Veya uses a hybrid thinking policy.

Reasoning, mathematics and planning examples use thinking blocks. Other responses use an empty thinking block.

Structured tool protocol

Veya can generate structured requests for an external agent runtime, including:

Tool calls
MCP tool calls
Skill requests
Memory requests
Clarification requests
Error responses

The model does not execute device actions by itself. A separate runtime must interpret requests, enforce permissions and execute approved actions.

GGUF downloads

Quantized GGUF files are being published in this repository.

File	Status
Veya-4B-Q4_K_M.gguf	Q4 quantization produced
Veya-4B-Q5_K_M.gguf	Pending
Veya-4B-Q8_0.gguf	Pending

Q4_K_M is approximately 2.33 GiB in the current export.

Usage

Use a compatible llama.cpp-based runtime to load a GGUF file.

Example:

llama-cli \
  -m Veya-4B-Q4_K_M.gguf \
  -p "Hello, introduce yourself."

Use the matching tokenizer and ChatML conversation format when integrating the model into an application.

Limitations

Responses can contain incorrect or outdated information.
Structured tool requests require a separate runtime.
Tool execution must be permission-controlled.
Android speed, RAM usage and thermal performance have not yet been benchmarked.
The small evaluation benchmark is indicative only.

Credits and attribution

Creator and fine-tuning: NotRIK

Base model: Qwen3-4B-Base, developed by the Qwen team at Alibaba Cloud.

This model is a fine-tune of the upstream Qwen model. It is not an independent model trained from scratch.

The upstream model's license and attribution requirements continue to apply.

License

Apache License 2.0, subject to the upstream model's license terms and notices.

Downloads last month: 88

GGUF

Model size

4B params

Architecture

qwen3

Hardware compatibility

4-bit

Model tree for NotRIK/Veya-4B-GGUF

Base model

Qwen/Qwen3-4B-Base

Quantized

(64)

this model