Chinese AI Unicorn's $34B LLM Surpasses Llama 2 and Falcon in Performance

01.AI Launches Yi-34B: A New Contender in Large Language Models

01.AI, a Chinese startup led by AI expert Kai-Fu Lee, has unveiled the Yi-34B, a large language model (LLM) boasting 34 billion parameters. This model surpasses competitors such as Meta's 70-billion parameter Llama 2 and the Technology Innovation Institute's 180-billion parameter Falcon.

The Yi-34B model is multilingual, supporting both Chinese and English, and can be customized for various applications. Additionally, a smaller model with 6 billion parameters is available, performing respectably on standard AI/ML benchmarks.

Expanding to Commercial AI Solutions

Launched just eight months ago, 01.AI has already achieved unicorn status and aims to intensify its offerings, planning a commercial product to compete with OpenAI, the current leader in generative AI by user numbers. This strategy aligns with a growing global trend where companies develop generative AI tailored to their specific markets.

Embracing the AI 2.0 Era

Founded in March, 01.AI seeks to usher in the AI 2.0 era, enhancing human productivity and driving significant economic and societal changes through advanced language models. The company emphasizes the transformative potential of AI 2.0, claiming it will create opportunities ten times larger than those of the mobile internet, reshaping software and user interfaces in the process.

Lee swiftly assembled a talented tech team featuring AI specialists from renowned firms like Google, Huawei, and Microsoft Research Asia. Initial funding came from Sinovation Ventures and Alibaba's cloud unit, while the exact funding amount remains undisclosed.

Performance Validation and Open Research Access

The initial release includes two bilingual models (6B and 34B parameters), both trained on sequences of 4,000 tokens, with the capacity to expand to 32,000 tokens during inference. The 34B model has demonstrated remarkable performance on platforms like Hugging Face, outshining its larger counterparts—achieving scores of 80.1 and 76.4 in common reasoning and reading comprehension tasks, compared to Llama 2's scores of 71.9 and 69.4.

The Yi series models offer an efficient solution for users, saving computational resources while allowing for cost-effective customization. Currently, the models are fully accessible for academic research, though commercial use requires explicit permissions.

Future Innovations on the Horizon

01.AI's models present appealing opportunities for organizations aiming to serve clients in China, facilitating the development of bilingual chatbots. The startup intends to broaden language support in its open-source models and is working on a commercial LLM to rival OpenAI’s GPT series, although details remain scarce.

01.AI is part of a broader movement among AI startups focused on localized LLMs. Baidu recently launched its ERNIE 4.0, showcasing new applications like Qingdu, a creative platform aiming to compete with Canva and Adobe Creative Cloud. Similarly, Korean company Naver is advancing with HyperCLOVA X, tailored for Korean language and cultural contexts, while India’s Reliance Industries collaborates with Nvidia to create an LLM encompassing the country's diverse languages for various applications.