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

Libraries

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

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

pipe = pipeline("text-generation", model="AuroraSystem/Velary-1.2-Flash-1.2B")
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-Flash-1.2B")
model = AutoModelForCausalLM.from_pretrained("AuroraSystem/Velary-1.2-Flash-1.2B", 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-Flash-1.2B 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-Flash-1.2B"
# 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-Flash-1.2B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

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

SGLang

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

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

Velary 1.2 Flash

English

Velary 1.2 Flash is a 1.17B-parameter bilingual (Russian/English) instruction model built on LiquidAI/LFM2.5-1.2B-Thinking (LFM2 architecture, 128K context) with switchable reasoning depth via a control tag appended to the user message.

Benchmarks (ABench75, 375 tasks, automatic regex scoring)

Category	none	medium	high
Overall	76.8%	82.7%	81.6%
Math	91%	97%	96%
Logic	69%	49%	43%
Code	71%	85%	85%
Cybersecurity	88%	96%	96%
General	65%	85%	88%
English	85%	84%	81%
Russian	65%	81%	83%

Training details

Base model: LiquidAI/LFM2.5-1.2B-Thinking (Lfm2ForCausalLM, 1.17B, vocab 65536, native thinking).
Method: LoRA, r=32, alpha=32, all-linear, dropout 0.05, merged into the base weights.
Data: approximately 8.7k generated domain examples (cybersecurity, code, math, logic, Russian language rules, identity, general) plus level data (none/medium/high) and a replay set for retention.
Stages: supervised fine-tuning on the domain pack (2 epochs), followed by two low-LR balancing stages (learning rate 3e-5) mixing level data, replay and a no-reasoning slice; adapters were merged between stages.
Precision and batching: fp16, batch size 1-2, gradient accumulation 8-32, maximum sequence length 1024-2048, gradient checkpointing.

Reasoning levels

You are Aurora, a helpful AI assistant. Respond in the user's language.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("AuroraSystem/Velary-1.2-Flash-1.2B")
model = AutoModelForCausalLM.from_pretrained(
    "AuroraSystem/Velary-1.2-Flash-1.2B", 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 + "\n<|effort=medium|>"},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

Limitations

Logic is the weakest category across all levels.
Casual conversation and open-ended creative writing are limited.
Longer reasoning increases token usage; set max_new_tokens of at least 512 for medium and high.
The model is not safety-aligned beyond the training data.

Русский

Velary 1.2 Flash — двуязычная (русский/английский) инструктивная модель на 1.17B параметров на базе LiquidAI/LFM2.5-1.2B-Thinking (архитектура LFM2, контекст 128K) с переключаемой глубиной рассуждений через управляющий тег в сообщении пользователя.

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

Категория	none	medium	high
Общий балл	76.8%	82.7%	81.6%
Математика	91%	97%	96%
Логика	69%	49%	43%
Код	71%	85%	85%
Кибербезопасность	88%	96%	96%
Общие знания	65%	85%	88%
Английский	85%	84%	81%
Русский	65%	81%	83%

Детали обучения

Базовая модель: LiquidAI/LFM2.5-1.2B-Thinking (Lfm2ForCausalLM, 1.17B, словарь 65536, нативный thinking).
Метод: LoRA, r=32, alpha=32, все линейные слои, dropout 0.05, с последующим мерджем в базовые веса.
Данные: около 8.7k сгенерированных доменных примеров (кибербезопасность, код, математика, логика, правила русского языка, идентичность, общие знания), плюс уровневые данные (none/medium/high) и replay-набор для сохранения навыков.
Этапы: SFT на доменном пакете (2 эпохи), затем два этапа балансировки с малым learning rate (3e-5) на смеси уровневых данных, replay и среза без рассуждений; адаптеры мерджились между этапами.
Точность и батчи: fp16, batch 1-2, накопление градиента 8-32, максимальная длина 1024-2048, gradient checkpointing.

Уровни рассуждений

You are Aurora, a helpful AI assistant. Respond in the user's language.

Ограничения

Логика — самая слабая категория на всех уровнях.
Слабо развита светская беседа и открытое креативное письмо.
Длинные рассуждения увеличивают расход токенов; для medium и high ставьте max_new_tokens не меньше 512.
Модель не выровнена по безопасности сверх обучающих данных.