AuroraSystem/Velary-1.0-1.2B

Aurora Velary 1.0

1.2B · LFM2 · 128K context · Development by: Aurora AI

---

Русская версия

Обзор

Компактная двуязычная (ru/en) генеративная модель с тремя режимами рассуждений. Веса — стандартный safetensors (bf16), совместим с transformers и llama.cpp (GGUF).

параметр	значение
Параметры	1.17B
Архитектура	LFM2 (hybrid conv + attention)
База	LiquidAI/LFM2.5-1.2B-Thinking
Контекст	128 000 токенов
Языки	русский, английский
Формат весов	safetensors (bf16)
Токенизатор	LFM2.5 (vocab 65 536)

Режимы рассуждений

Три уровня глубины: none, medium, high.

уровень	поведение
none	прямой ответ, без видимого рассуждения
medium	краткое рассуждение, затем ответ
high	развёрнутое рассуждение, затем ответ

Уровень задаётся тегом в конце пользовательского сообщения:

System: You are Aurora, a helpful AI assistant. Respond in the user's language.
User:   <вопрос>
        <|effort=high|>

ABench75

Собственный авто-проверяемый комплект: 375 задач, 75 на категорию (math / logic / code / cyber / general), английский + русский.

режим	overall
none	69.6%
medium	75.5%
high	77.1%

категория	LFM2.5-1.2B	Qwen3-0.6B	Qwen2.5-1.5B	Qwen3-1.7B	Velary 1.0 high
Math	44%	91%	85%	89%	93%
Cyber	88%	84%	93%	97%	91%
Russian	52%	60%	85%	78%	73%

Использование

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("AuroraAI/velary-1.0")
model = AutoModelForCausalLM.from_pretrained("AuroraAI/velary-1.0", 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": "Explain how binary search works.\n<|effort=high|>"},
]
ids = tok.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

llama.cpp (GGUF):

llama-server -m velary-1.0-f16.gguf --jinja -c 131072 -ngl 99

Ограничения

• Многошаговое математическое reasoning (GSM8K) — ниже 1.5B-моделей (~56% zero-shot).
• Логические задачи — слабое место (~51%).
• Русский уступает Qwen3-1.7B/Qwen2.5-1.5B (база LFM2.5 англоцентрична).

---

English version

Overview

A compact bilingual (ru/en) generative model with three reasoning modes. Weights are standard safetensors (bf16), compatible with transformers and llama.cpp (GGUF).

parameter	value
Parameters	1.17B
Architecture	LFM2 (hybrid conv + attention)
Base	LiquidAI/LFM2.5-1.2B-Thinking
Context	128,000 tokens
Languages	Russian, English
Weight format	safetensors (bf16)
Tokenizer	LFM2.5 (vocab 65,536)

Reasoning modes

Three depth levels: none, medium, high.

level	behavior
none	direct answer, no visible reasoning
medium	brief reasoning, then answer
high	extended reasoning, then answer

The level is set by a tag at the end of the user message:

System: You are Aurora, a helpful AI assistant. Respond in the user's language.
User:   <question>
        <|effort=high|>

ABench75

In-house auto-graded suite: 375 tasks, 75 per category (math / logic / code / cyber / general), English + Russian.

mode	overall
none	69.6%
medium	75.5%
high	77.1%

category	LFM2.5-1.2B	Qwen3-0.6B	Qwen2.5-1.5B	Qwen3-1.7B	Velary 1.0 high
Math	44%	91%	85%	89%	93%
Cyber	88%	84%	93%	97%	91%
Russian	52%	60%	85%	78%	73%

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("AuroraAI/velary-1.0")
model = AutoModelForCausalLM.from_pretrained("AuroraAI/velary-1.0", 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": "Explain how binary search works.\n<|effort=high|>"},
]
ids = tok.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

llama.cpp (GGUF):

llama-server -m velary-1.0-f16.gguf --jinja -c 131072 -ngl 99

Limitations

• Multi-step math reasoning (GSM8K) is below 1.5B models (~56% zero-shot).
• Logic tasks are a weak spot (~51%).
• Russian trails Qwen3-1.7B / Qwen2.5-1.5B (the LFM2.5 base is English-centric).