DeepSeek V4 Goes Live - Open-Weight 1.6T MoE and 1 Million Token Context Window

DeepSeek has released V4 Preview, delivering two open-weight models that push the boundaries of accessible AI capabilities. The V4-Pro features 1.6 trillion parameters with mixture-of-experts architecture activating 49 billion parameters per token, whilst the V4-Flash variant operates with 284 billion total parameters and 13 billion active parameters. Both models support a 1 million token context window, marking a significant advancement in long-context processing. The Chinese laboratory, which previously disrupted the industry with its V3 and R1 models, claims these new variants achieve parity with leading closed-source models whilst operating at substantially lower costs. The models are immediately accessible through chat.deepseek.com via new Expert Mode and Instant Mode toggles, with API access also updated to support the new capabilities. In keeping with DeepSeek's commitment to open research, the model weights have been published on HuggingFace alongside a comprehensive technical report. This immediate availability enables the broader research community to examine and build upon the technology without delay. The unannounced Friday release of a trillion-parameter open-weight model with million-token context represents another significant shift in AI accessibility. As frontier capabilities become increasingly available through open models, the competitive advantages of closed-source approaches continue to diminish.