We hereby present the first series of small open models by Aqui, trained from scratch and with an MIT license. The open1 family delivers state-of-the-art performance across reasoning, mathematics, and coding tasks while maintaining efficient inference capabilities.
Ultra-efficient model optimized for edge deployment and real-time applications.
Balanced model providing exceptional performance across diverse tasks with reasonable compute requirements.
Aqui-open1-4x8B — Our biggest non-thinking open model, head-to-head against Qwen3 32B and Llama 3.3 70B. Stay tuned for the most capable open model in the series.
| Metric | open1-1.5B-Instruct | Llama-3.2-1B-Instruct | LFM2-1.2B | Qwen3-1.7B | Gemma-3-1B-it | SmolLM2-1.7B-Instruct |
|---|---|---|---|---|---|---|
| MMLU | 58.5 | 46.6 | 55.2 | 59.1 | 40.1 | 42.3 |
| GPQA | 32.3 | 28.8 | 31.5 | 27.7 | 21.1 | 22.1 |
| GSM8K | 62.6 | 35.7 | 58.3 | 51.4 | 59.6 | 48.2 |
| IFEval | 72.7 | 52.4 | 74.9 | 74.0 | 62.9 | 56.7 |
| MGSM | 59.1 | 29.1 | 55.0 | 66.6 | 43.6 | 38.5 |
| Average | 57.0 | 38.5 | 55.0 | 55.8 | 45.5 | 41.6 |
| Benchmark | open1-7.5B-Instruct | Llama-3.1-8B-Instruct | LFM-7B | Qwen3-8B | Gemma-3-12B-it | Nemotron-Nano-9B-v2 |
|---|---|---|---|---|---|---|
| MMLU | 75.8 | 68.7 | 69.4 | 71.6 | 72.5 | 74.5 |
| HumanEval | 82.3 | 71.7 | 70.1 | 84.8 | 84.8 | 86.2 |
| GPQA Diamond | 52.2 | 25.9 | 32.9 | 45.2 | 34.9 | 40.8 |
| IFEval | 78.9 | 77.0 | 71.6 | 83.4 | 81.5 | 84.3 |
| AIME 2025 | 18.9 | 4.3 | 2.1 | 20.2 | 18.3 | 20.1 |
| Average | 61.6 | 49.5 | 49.2 | 61.0 | 58.4 | 61.2 |
Exceptional performance on MMLU, GPQA, and mathematical reasoning tasks, outperforming models of similar and larger sizes.
Efficient transformer design enabling fast inference while maintaining high accuracy across diverse benchmarks.
Trained on 20+ languages with robust performance across linguistic boundaries and cultural contexts.
Complete freedom for commercial use, modification, and redistribution with minimal restrictions.
from transformers import AutoTokenizer, AutoModelForCausalLM
# Load the model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("aquigpt/open1-1.5B-Instruct")
model = AutoModelForCausalLM.from_pretrained("aquigpt/open1-1.5B-Instruct")
# Generate text
inputs = tokenizer("Explain quantum computing:", return_tensors="pt")
outputs = model.generate(**inputs, max_length=200, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
The open1 models were trained from scratch on a diverse, high-quality dataset spanning code, mathematics, reasoning, and multilingual text. Training utilized advanced techniques including:
Note: These models are designed for research and commercial applications. While they demonstrate strong performance, users should conduct appropriate testing for their specific use cases.
Built with ❤️ by the Aqui team • MIT • September 2025