DeepSeek, a Chinese AI startup, has unveiled its latest large language models, DeepSeek V4 Flash and DeepSeek V4 Pro, which feature context windows of over 1 million tokens. The V4 Pro model boasts an impressive 1.6 trillion parameters, making it the largest open-weight model available, surpassing competitors like Moonshot AI's Kimi K 2.6. The V4 Flash model, with 284 billion parameters, is designed for efficiency, allowing users to input large documents or codebases. Both models are released under research preview and are focused solely on text outputs, unlike many closed rivals that can generate audio and video. DeepSeek's architectural advancements, including mixture-of-experts (MoE) and multi-head latent attention (MLA), enhance performance while keeping costs lower than leading models. The V4 Flash model is priced at $0.14 per million input tokens, while the V4 Pro model costs $0.145 per million input tokens, both undercutting competitors like OpenAI and Google. Despite trailing slightly behind the latest frontier models in knowledge tests, DeepSeek's new offerings challenge existing assumptions about AI training costs and performance.