Chinese AGI Startup ‘StepFun’ Developed ‘Step-2’: A New Trillion-Parameter MoE Architecture Model Ranking 5th on Livebench

In the evolving landscape of artificial intelligence, building language models capable of replicating human understanding and reasoning remains a significant challenge. One major hurdle in the development of large language models (LLMs) is balancing computational efficiency with expansive capabilities. As models grow larger to capture more complex relationships and generate better predictions, the computational costs increase significantly. Meanwhile, general-purpose LLMs must handle a range of tasks—such as instruction following, coding, and reasoning—often struggling to maintain consistent performance across all dimensions. This inconsistency poses a notable bottleneck, particularly for those aiming to advance toward artificial general intelligence (AGI).

Introducing Step-2: A Trillion-Parameter MoE Model

StepFun, a Shanghai-based AI startup focused on advancing AGI, has recently developed Step-2, a trillion-parameter Mixture of Experts (MoE) language model. This model has gained attention by ranking 5th on Livebench, a prominent global benchmarking platform that evaluates AI models based on their overall performance across diverse tasks. Step-2 is the first trillion-parameter MoE model developed by a Chinese company and ranks as China’s top-performing LLM. It holds its position behind some of the most advanced models from industry leaders like OpenAI and Google. This achievement reflects the advanced technology StepFun is building and its effort to contribute to the global AI community from within China.

Architecture and Technical Insights

The Step-2-16k model is built using MoE architecture, a design approach that allocates computational resources more efficiently compared to traditional fully-dense models. Mixture of Experts uses a routing mechanism that activates only a subset of the model’s parameters—the experts—for any given task, enabling the scaling of parameters without proportionally increasing computation. The trillion-parameter scale allows Step-2 to capture a nuanced understanding of language, offering substantial improvements in instruction-following capabilities and reasoning tasks. It also supports a context length of up to 16,000 tokens, which is particularly useful for applications requiring long-term dependencies, such as document analysis or complex conversations.

Performance Metrics and Areas for Improvement

Technically, the Step-2 model has demonstrated a range of strengths, with high scores in several areas. The model achieved an Instruction Following (IF) score of 86.57, indicating its ability to comprehend and act upon complex instructions. Additionally, Step-2 secured a reasoning score of 58.67 and a data analysis score of 54.86, highlighting its proficiency in processing and understanding information. However, the model showed room for improvement in coding and mathematics, scoring 46.87 and 48.88, respectively. Despite these areas needing further optimization, Step-2 effectively leverages MoE to balance parameter scale with task-specific efficiency. The model’s development focused heavily on research and development (R&D) rather than marketing, ensuring robust performance and reliability even at this large scale.

Significance and Accessibility

The significance of Step-2 lies in both its scale and its competitive edge as the first trillion-parameter model from a Chinese startup to achieve such a high ranking. As the AI community grows increasingly concerned with accessibility and inclusiveness, StepFun has made Step-2 accessible through its API platform, making it available for developers and researchers. Additionally, Step-2 has been integrated into the consumer application “Yuewen,” broadening its reach and offering the general public an opportunity to interact with a state-of-the-art language model. The model’s ranking—5th globally—demonstrates that Chinese startups are capable of producing high-quality AI systems, and it suggests a future where diverse players contribute significantly to the AI field, thereby reducing the concentration of AI expertise among only a few Western companies.

Conclusion

StepFun’s Step-2 represents progress not only for the company but also for the Chinese AI community. By ranking 5th on Livebench, Step-2 showcases its capability in areas like instruction following and reasoning, while also highlighting areas where further refinement is needed, such as coding and mathematics. Built with an MoE architecture and equipped with a trillion parameters, Step-2’s strengths are a testament to the thoughtful application of advanced architectures for creating expansive and efficient models. With its accessible implementation via APIs and consumer integration, Step-2 also demonstrates StepFun’s commitment to bringing advanced technology to users worldwide. While there is work to be done, particularly in enhancing coding and mathematical capabilities, Step-2’s performance and architecture signify the increasing maturity of AI research and development from regions beyond the traditional powerhouses. This accomplishment positions StepFun as a key player in the AI landscape, setting the stage for further developments in AGI research and industry applications.

---

Check out the Details. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. If you like our work, you will love our newsletter.. Don’t Forget to join our 55k+ ML SubReddit.

[FREE AI VIRTUAL CONFERENCE] SmallCon: Free Virtual GenAI Conference ft. Meta, Mistral, Salesforce, Harvey AI & more. Join us on Dec 11th for this free virtual event to learn what it takes to build big with small models from AI trailblazers like Meta, Mistral AI, Salesforce, Harvey AI, Upstage, Nubank, Nvidia, Hugging Face, and more.