DeepSeek Official Blog Introduces Features of DeepSeek V4

According to Golden Finance, on April 10, the official DeepSeek blog published an article introducing DeepSeek V4. DeepSeek V4 is DeepSeek's upcoming flagship model, which not only pushes the limits of parameter scale but also promises unprecedented efficiency. DeepSeek V4 is expected to handle 1 trillion (1T) parameters, natively support multimodal data such as text, images, videos, and audio, and feature a context window of 1 million tokens (equivalent to 15–20 complete novels), making it a direct competitor to Western giants like OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.5. The API pricing for DeepSeek V4 is 10–50 times cheaper than GPT-5.4 and Claude Opus 4.5. It is anticipated that DeepSeek V4 will adopt the Apache 2.0 license for open-sourcing its weights. DeepSeek V4 can run locally on dual RTX 4090 GPUs or a single RTX 5090 GPU. DeepSeek also introduced three groundbreaking innovations in DeepSeek V4: 1. Engram Conditional Memory 2. Manifold-Constrained Hyperconnection (mHC) 3. Sparse Attention Mechanism (DSA) and Lightning Indexer Additionally, DeepSeek stated that due to strict U.S. export restrictions on high-end NVIDIA GPUs (such as B300 and H200), DeepSeek has optimized V4 to primarily rely on Chinese domestic chips for inference. While initial training may still utilize NVIDIA hardware (such as H800s), the model has been highly optimized for Huawei Ascend 950PR and Cambricon MLU chips.