GuShiio.com 
Step Star’s latest open source video model Step-Video-T2V effect (Source: Provided by respondents)
The craze for large open source models triggered by DeepSeek continues. Today, more than five AI models have announced open source news.
GuShiio.comAGI获悉，2月18日上午，“Step Star of General Artificial Intelligence Company, one of the six tigers of the big model, and Geely Automobile Group jointly announced that they will open source the two Step series multimodal large models of the two collaborative Step series to developers around the world.
Among them,Step-Video-T2V, the world’s largest parameter and best performance open source video generation model, will be open source, with a parameter volume of 30 billion, and can directly generate high-quality video with 204 frames and 540P resolution.
Step Star’s second open source model is the industry’s first product-level open source voice interaction model Step Audio. It has a parameter scale of up to 130 billion. It can generate emotions, dialects, languages, singing and personalization according to different scene requirements. Style expression can naturally have high-quality dialogue with users, high-quality sound reproduction and role-playing, meeting application needs in film and television entertainment, social networking, games and other industry scenarios.
In mainstream public test sets, the Step Step-Audio model performs well. In addition, Step Star has also built and open-source the multi-dimensional evaluation system StepEval-Audio-360 benchmark.
Step Star’s third open source model is a new benchmark dataset Step-Video-T2V-Eval for Wensheng video quality evaluation. It contains 128 Chinese evaluation questions from real users and aims to evaluate the quality of generated videos in 11 content dimensions such as sports, scenery, animals, combination concepts, surreality, characters, 3D animation, and cinematography. The evaluation results show that the model performance of the Step Step-Video-T2V is excellent in terms of command compliance, motion smoothness, physical rationality, and aesthetics.
Currently, you can experience the video generation capabilities of Step Step-Video-T2V on both the Yuewen website and Yuewen App. It is worth noting that the earlier Step Star Leaking Product was officially connected to the DeepSeek model to provide in-depth thinking services.
It’s not just a step up star. On the morning of February 18,Kunlun Wanwei announced that it has jointly released SkyReels-V1, the first video generation model for AI short drama creation, and China’s first SOTA-level emotion-action controllable algorithm SkyReels-A1 based on a video dock model.
Among them, Kunlun Wanwei said that SkyReels-V1 not only supports Wensheng videos, but also graphics videos. It is the model with the largest parameter among the open source video generation models to support graphics videos. All indicators can achieve open source SOTA at the same resolution.
SkyReels-V1 comparison of cultural and student video indicators
At the computing power level, Kunlun Wanwei said that with the support of the self-developed reasoning optimization framework SkyReels-Infer, V1 has greatly improved reasoning efficiency and achieved 544p resolution. The reasoning is based on a single 4090 in only 80 seconds, and it also supports distributed multi-card parallelism. With the same RTX4090 resources, the SkyReels-Infer version reduces the end-to-end latency by 58.3%(293.3s vs 464.3s) compared with the official version of Tencent Hunyuan Video. In addition, new technologies can be adopted to meet the operating requirements of user-level graphics cards with low memory, and support model compilation optimization to further optimize latency. Based on the open source diffuser library, ease of use can be improved.
Kunlun Wanwei said that the open source of SOTA-level SkyReels-V1 and SkyReels-A1 at the same time is the first case in the AI short drama industry. It is also a small step for Kunlun Wanwei SkyReels to give back to the industry. It is also a step towards promoting AI short drama creation and video generation. A big step towards the flourishing of the industry. In the future, the cross-border development of short dramas and games, virtual reality and other fields will accelerate industrial integration, and AI short dramas also have the hope of moving from technical experiments to mainstream creation and becoming a new carrier of global cultural output.
In fact, since mid-January, DeepSeek, a China open source AI model, has been born, shaking the entire AI technology industry. Its cost is cheap. DeepSeek V3 completed the training in just two months at a cost of US$5.6 million, which is only a small part of the amount spent by companies such as OpenAI. On the other hand, DeepSeek is an open source model that quickly attracted the intervention of users from Internet technology companies and other fields.Especially for computing power and AI talents, it will have a new promoting effect.
On February 16, the “China Artificial Intelligence Computing Power Development Assessment Report” jointly released by IDC and Inspur Information showed that in 2024, the scale and market size of China’s intelligent computing power will increase by 74.1% and 86.9% respectively year-on-year. It is expected that in 2025, China’s intelligent computing power scale will increase by 43% compared with 2024; the scale of China’s artificial intelligence computing power market will reach US$25.9 billion, an increase of 36.2% compared with 2024.
GuShiio.comAGI learned from the Enterprise Inspection Office that as of now, there are 647 existing computing-related companies in China. In the past ten years, the number of registered enterprises has shown an overall growth trend. In 2024, 207 related enterprises will be registered throughout the year, a year-on-year increase of 52.21%. In 2025, 15 AI-related enterprises have been registered in China.
Judging from the distribution of registered capital, more than 40% of the relevant enterprises with a registered capital of more than 10 million yuan. From an industry perspective, more than 40% of the relevant enterprises belong to the scientific research and technical service industries.
At the talent level,The latest report released by Zhaopin Recruitment shows that from the perspective of job seekers, the growth rate of job seekers in the computer hardware and computer software industries in the second week was 49.9% and 38.6% respectively, ranking among the top two in the industry. Job seekers in IT services, communications/telecommunications/network equipment industries also increased by 30% month-on-month; from the perspective of occupation, job seekers in technical positions such as front-end development, software research and development, mobile R & D, test engineers, artificial intelligence engineers, communications and hardware R & D are all 30%-50% month-on-month; In terms of recruitment salary, in the second week after the holiday, the average monthly recruitment salary in the computer software and computer hardware industries was 11360 yuan and 10660 yuan respectively, up 8.3% and 5.9% respectively from the first week.
The report pointed out that AI development has a boost effect on the entire information technology industry, with the supply and demand of computer hardware/software talents and wages rising.
Just on February 18, OpenAI CEO Sam Altman said that open-source next-generation model and asked everyone what kind of open source project they would like to build an o3-mini model that is quite small but still needs to run on a GPU., or the best mobile phone-sized model you can make.“”“”
This means that OpenAI is about to open source a large model, and this move is undoubtedly a positive response to the current trend of AI open source.
The same day that Ultraman posted the article was the time Musk announced that he would release the world’s smartest artificial intelligence. Musk’s artificial intelligence startup xAI will release the latest version of the Grok 3 chat robot. Although it is several months later than originally planned, it still attracts a lot of attention.
