SuperAPI AI Directory
Partner with us中文AI newsFavorites

All AI tools

35,708 results · Page 877 of 992

station

AI agents

Station is an open-source multi-agent environment that simulates a miniature scientific ecosystem, allowing AI agents to autonomously conduct scientific discovery and jointly build a shared literature. It is suitable for scorable, fast-iterating (completed in about 2 hours) research tasks, such as architecture search, code discovery, optimization, computational biology, mathematical construction, and data analysis. The project is maintained by dualverse-ai and provides multiple versions such as v1, v1.5, and v2, and has made new discoveries on mathematical tasks. Use requires an API key from a model provider and the OpenAI Codex CLI, and supports human-machine interaction, backup and branching, custom research tasks, and more through a dashboard.

Read more

Steel-LLM

AI developer platforms

Steel-LLM is an open-source project initiated by an individual to pretrain a Chinese large model from scratch, using 1T tokens of data to train a Chinese LLM with approximately 1B parameters. From the start of the project to fine-tuning the first version took about 8 months. It shares in detail the entire process of data collection, data processing, pretraining framework selection, model design, and more, and has open-sourced all code, with the goal of being reproducible under conditions of having 8 to several dozen GPUs. Its name "Steel" is inspired by the band Omnipotent Youth Society, implying "backyard steelmaking" under limited conditions. According to the official introduction, thanks to open-source Chinese data, Steel-LLM performs better on Chinese benchmarks than some larger models released early by institutions, reaching 42 points on CEVAL and 36 points on CMMLU. The project also includes pretraining programs, data preparation scripts, fine-tuning and evaluation details, and has released a technical report and related blogs. It will continue to explore mathematical capabilities, reinforcement learning, complex reasoning, and other areas in the future.

Read more

Step-Video-T2V

AI video

Step-Video-T2V is an open-source text-to-video pretrained model from stepfun-ai with 30 billion parameters, capable of generating videos up to 204 frames long. The model uses a deeply compressed Video-VAE (16x16 spatial and 8x temporal compression ratios) to improve training and inference efficiency, and uses bilingual text encoders to support Chinese and English prompts. It is based on the DiT architecture (48 layers, 48 attention heads) combined with 3D full attention and Flow Matching for denoising generation, and finally improves visual quality through Direct Preference Optimization (DPO). The project released inference code and model weights on February 17, 2025, along with a Turbo version; on March 17, 2025, it released Step-Video-TI2V, an image-to-video model based on this model. The model needs to run on NVIDIA GPUs that support CUDA, and an 80GB VRAM GPU is recommended. The test environment is Linux, and it depends on Python >= 3.10.0 and PyTorch >= 2.3-cu121. The official team also released Step-Video-T2V-Eval, an evaluation benchmark containing 128 Chinese prompts, covering 11 categories. The online version can be experienced in Yuewen Video.

Read more