The exciting new leak of DeepSeek V4!

✅ 1 million token long context: That's right, this is real! Version 4 will support a 1 million token context window. This means it can handle ultra-long texts like the trilogy of "The Three-Body Problem" in one go, or process extensive codebases and perform deep logical reasoning.
🎯 Key upgrade highlights:
Native multimodal capabilities: Not only supports text but can also understand images, enabling combined text and image analysis;
Trillions of parameters: Expected to be a giant foundational model with trillions of parameters;
Enhanced programming abilities: Significantly optimized for multi-file project understanding and long chain reasoning;
Stunning SVG generation: Test samples show that V4's SVG graphics quality far surpasses the previous generation, capable of outperforming V3.2 even in non-thinking mode;
⏰ Release date:
According to Reuters, DeepSeek plans to release V4 as early as next week. Due to the significant increase in model size, training speed has slowed, and the release has been delayed from the original schedule.
😊 An interesting strategic shift:
This time, DeepSeek broke industry norms by prioritizing opening testing access to Chinese chip manufacturers like Huawei, while NVIDIA and AMD were left on the sidelines. This is seen as an important signal to strengthen the domestic computing power ecosystem.