Community Open Source Implementation of GPT4o in PyTorch

• TikToken Tokenzier: We know fursure the tokenizer. Which is here
• Model understands Images and Audio Natively. There are 2 approaches, process them natively or use encoders for each. I think here they're using encoders like whisper and vit for simplicity and brevity.
• Using DALLE3 as the output head to generate images
• Tokens to denote when to generate an image or audio
• Whisper output head for the audio outputs

MIT