Intern: SenseTime's General-Purpose Foundation Model 134

Shusheng General Large Model is a large pre-trained model released by the Shanghai Artificial Intelligence Laboratory.

Shusheng General Large Model Products:

Shusheng · Multimodal Large Model

Shusheng 2.5 is a version of the multimodal multi-task general large model, jointly released by SenseTime and the Shanghai Artificial Intelligence Laboratory. Trained on 8 billion multimodal samples with 20 billion parameters. It has broken through multiple key technologies such as cursor instruction interaction, defining arbitrary tasks using language, and lightweight adaptive fusion, achieving three major capabilities: open-world understanding, multimodal interaction, and cross-modal generation, supporting 3.5 million semantic tags.

Outstanding capabilities: open understanding, multimodal interaction, cross-modal generation

Original technologies: defining arbitrary tasks using language, cursor instruction interaction, lightweight adaptive fusion

Shusheng · Puyu InternLM

A base model with hundreds of billions of parameters. Through multi-stage progressive training, it possesses a high level of knowledge and strong reasoning capabilities, performing outstandingly in comprehensive exams and all-around capability evaluations.

InternLM is a multilingual base model with hundreds of billions of parameters trained on over a trillion tokens of data. Through multi-stage progressive training, the InternLM base model has a high level of knowledge and performs excellently in scenarios requiring strong thinking abilities such as Chinese and English reading comprehension and reasoning tasks, and stands out in various comprehensive exams designed for humans. On this basis, by combining high-quality human-labeled dialogue data with techniques such as RLHF, InternLM can respond to complex instructions when conversing with humans and produce replies that align with human morality and values.

Shusheng · Horizon - Real-Scene 3D Large Model

An internationally leading city-level real-scene 3D large model based on NeRF, with extremely high training and rendering efficiency.

Large-Scale, High-Precision Modeling:For the first time, it achieves efficient training of a city-level NeRF with 100 square kilometers and 4K image precision, and the modeling range is theoretically infinitely scalable.

Real-Time, High-Precision Rendering:For the first time, it achieves 1k resolution, 30 frames real-time rendering across the full 100 square kilometer range of city-level NeRF, as well as 4k resolution offline rendering, delivering an immersive shuttle experience.

Rich Extended Features:Provides a wealth of capabilities beyond reconstruction, including city editing features such as removing, creating, and rotating urban buildings, and city style adjustment features such as altering lighting and seasons.

Integrated Training, Rendering, and Interaction System:For the first time, it implements a city-level NeRF system covering algorithms, operators, computing systems, and user interaction, providing a solid foundation for training, rendering, interaction, and applications.