Instructions to use AuroraSystem/Velary-1.2-Mini-0.35B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AuroraSystem/Velary-1.2-Mini-0.35B with Transformers:

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

pipe = pipeline("text-generation", model="AuroraSystem/Velary-1.2-Mini-0.35B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("AuroraSystem/Velary-1.2-Mini-0.35B")
model = AutoModelForCausalLM.from_pretrained("AuroraSystem/Velary-1.2-Mini-0.35B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AuroraSystem/Velary-1.2-Mini-0.35B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AuroraSystem/Velary-1.2-Mini-0.35B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AuroraSystem/Velary-1.2-Mini-0.35B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AuroraSystem/Velary-1.2-Mini-0.35B

SGLang

How to use AuroraSystem/Velary-1.2-Mini-0.35B 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 "AuroraSystem/Velary-1.2-Mini-0.35B" \
    --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": "AuroraSystem/Velary-1.2-Mini-0.35B",
		"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 "AuroraSystem/Velary-1.2-Mini-0.35B" \
        --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": "AuroraSystem/Velary-1.2-Mini-0.35B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AuroraSystem/Velary-1.2-Mini-0.35B with Docker Model Runner:
```
docker model run hf.co/AuroraSystem/Velary-1.2-Mini-0.35B
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Velary Mini

English

Velary Mini is a 0.35B-parameter bilingual (Russian/English) instruction model built on LiquidAI/LFM2.5-350M (LFM2 architecture, 128K context). It answers directly without a visible reasoning block.

Benchmarks (ABench75, 375 tasks, automatic regex scoring)

Category	Score
Overall	50.7%
Math	32%
Logic	39%
Code	67%
Cybersecurity	55%
General	61%
English	52%
Russian	48%

The model has a single direct-answer mode; there is no effort level.

Training details

Base model: LiquidAI/LFM2.5-350M (Lfm2ForCausalLM, about 0.35B, vocab 65536).
Method: LoRA, r=32, alpha=32, all-linear, dropout 0.05, merged into the base weights.
Data: about 9.7k examples assembled from the domain pack and the level mix; every record was converted to direct answers (no reasoning field, effort tag removed).
Precision and batching: fp16, batch size 2 with gradient accumulation 16, maximum sequence length 2048, gradient checkpointing, 2 epochs, learning rate 2e-4.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("AuroraSystem/Velary-1.2-Mini-0.35B")
model = AutoModelForCausalLM.from_pretrained(
    "AuroraSystem/Velary-1.2-Mini-0.35B", dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "You are Aurora, a helpful AI assistant. Respond in the user's language."},
    {"role": "user", "content": question},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

Limitations

Limited capacity for math and multi-step logic (see benchmark table).
Short, direct answers; no chain-of-thought.
The model is not safety-aligned beyond the training data.

Русский

Velary Mini — двуязычная (русский/английский) инструктивная модель на 0.35B параметров на базе LiquidAI/LFM2.5-350M (архитектура LFM2, контекст 128K). Отвечает напрямую, без видимого блока рассуждений.

Бенчмарки (ABench75, 375 задач, автоматическая проверка по регулярке)

Категория	Балл
Общий балл	50.7%
Математика	32%
Логика	39%
Код	67%
Кибербезопасность	55%
Общие знания	61%
Английский	52%
Русский	48%

У модели один режим — прямой ответ; уровней нет.

Детали обучения

Базовая модель: LiquidAI/LFM2.5-350M (Lfm2ForCausalLM, около 0.35B, словарь 65536).
Метод: LoRA, r=32, alpha=32, все линейные слои, dropout 0.05, с мерджем в базовые веса.
Данные: около 9.7k примеров из доменного пакета и уровневого микса; каждая запись приведена к прямому ответу (поле reasoning убрано, тег уровня удалён).
Точность и батчи: fp16, batch 2 с накоплением градиента 16, максимальная длина 2048, gradient checkpointing, 2 эпохи, learning rate 2e-4.