Stepping Stars debuts “Trillion” and “Multi-Mode” large models, and also has a cross-border linkage with “Havoc in Heaven” | WAIC 2024

In just the past 100 days or so, the stars standing at the forefront of AGI have once again shown great results.

At the World Artificial Intelligence Conference unveiled today, Step Star announced a comprehensive upgrade to the Step series of general-purpose large model families.

Step Star launches three new Step series general-purpose large models in one go: the official version of the Step-2 trillion parameter language large model, the Step-1.5V multi-modal large model, and the Step-1X large image generation model.

▲ APPSO is on site

The newly released official version of the Step-2 trillion parameter language model adopts the innovative MoE hybrid expert model architecture.

At the same time, based on Step Star's industry-leading system capabilities, Step-2 has greatly improved training efficiency, and its somatosensory experience is fully close to GPT-4 in terms of mathematics, logic, programming, knowledge, creation, and multiple rounds of dialogue. Enterprises and developers can Step Star Open Platform application experience.

In addition, Step Star also announced a number of progress in the research and development of multi-modal large models.

The newly upgraded Step-1.5V 100 billion parameter multi-modal large model has comprehensively improved image perception and understanding capabilities, and has excellent video understanding capabilities. It can accurately identify objects, people and environments in videos, and understand the overall atmosphere and character emotions of the video.

With the blessing of the Step-2 trillion-parameter large model, the reasoning capabilities of Step-1.5V have been greatly enhanced, and it can perform advanced reasoning tasks such as answering mathematical questions, writing code, and composing poetry based on image content.

If the release of Step-1.5V marks that Step Star has achieved a leapfrog upgrade from image understanding to video understanding in a very short period of time, then the newly released Step-1X image generation large model represents a step forward. Xingchen has made important progress in promoting a unified technical route for multi-modal understanding and generation.

It adopts the full-link self-developed DiT (Diffusion Models with transformer) model architecture and supports three different parameter quantities of 600M, 2B, and 8B, which can meet the needs of different scenarios.

In addition, Step-1X has more powerful semantic alignment and command following capabilities, and is deeply optimized for Chinese elements and culture, making it more Chinese-style. This is also an important feature that distinguishes Step-1X from other models.

It is worth mentioning that Step Star also won the title of WAIC 2024 SAIL Star with the innovation of the Step series of universal large models.

Dr. Jiang Daxin, founder and CEO of Step Star, once said:

To climb the AGI mountain, "trillions of parameters" and "multi-mode fusion" are indispensable. The scale of trillions of parameters is the basic threshold for realizing AGI; multi-modal large models are the only way to achieve AGI.

Facing the future, we will continue to make the model bigger and stronger, create a super model, and at the same time let the model serve work and life, tenfold everyone's possibilities.

Step Star also demonstrated its self-developed products for C-end users at WAIC, as well as its latest achievements and plans for application implementation.

Intelligent assistant – "Yuewen" and AI open world platform – "Bubble Duck" are two self-developed products launched by Step Star for C-end users.

Based on the powerful capability of the Step series of general large models, "Yuewen" can accurately describe and understand text, data, charts and other information in images, and excellently complete tasks such as content creation, logical reasoning, and data analysis.

"Bubble Duck" creates a brand new AI open world. Here, users can explore stories, create characters, and immerse themselves in their own open world.

At the same time, Step Star has reached in-depth cooperation with many partners to jointly explore innovative applications for C-end users.

For example, Step Star has reached an in-depth cooperation with Jiemian Financial Association, a subsidiary of Shanghai Media Industry, and the two parties have promoted the application of large models in the fields of AIGC financial information, intelligent investment research, and intelligent investment consulting.

APPSO is on site

Facing the field of content creation, Step Star has teamed up with Shanghai Film to conduct innovative explorations in the field of "AI+IP". The two parties launched an AI interactive experience for "Havoc in Heaven" – "Test Which God You Are", which uses the Step series of large models and integrates multiple capabilities such as image understanding, style transfer, image generation, and plot creation.

Step Star also worked with Shanghai Film to demonstrate its video generation capabilities using the IP of "Calabash Brothers". In addition, Step Star has cooperated with Chinese Online, China National Knowledge Infrastructure, etc. to jointly explore innovative applications of large models in online literature, knowledge services and other fields.

In addition, Step Star announced the launch of an open platform – "Star Project", investing high-quality resources to support outstanding start-ups and independent developer teams in the multi-modal field, and jointly accelerating the application of AI large model technology in various scenarios.

What few people know is that the name of step stars actually comes from the "step function".

This seems to coincide with the core concept of "Scaling Law" mentioned above. As the scale of the model continues to expand, its performance will also be significantly improved, and it will also usher in step progress.

The new Step series of general-purpose large models released today is only a small step on the road of exploration for Step Stars, but it is also an important step towards the road to AGI and using the power of Step to reach the sea of ​​stars.

In the future battlefield of domestic large-scale models, perhaps no one dares to ignore the dark horse of domestic large-scale models, Step Star.

# Welcome to follow the official WeChat public account of Aifaner: Aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo