AI video · SuperAPI AI Directory

Targum video

AI-powered video translation service for quick and easy language translation.

TeleTron

TeleTron is an open-source project developed by the Tele-AI team, aiming to advance the training of long-context multimodal Transformer models. Through hybrid parallelism strategies, activation checkpointing, and fused CUDA kernels, it optimizes GPU memory usage and supports training video clips longer than 30 seconds at 720P resolution. The project has released full-parameter training and inference code for HunyuanVideo, and supports Wan-2.1 i2v and Wan-2.2 training. Through a distributed multimodal encoder, it achieves over 30% acceleration in HunyuanVideo training. TeleTron features flexible parallelism strategies, supporting data parallelism, context parallelism, and tensor parallelism, making it suitable for large-scale and efficient training.