DeepSeek previews new AI model that ‘closes the gap’ with frontier models

DeepSeek released two preview versions of its V4 large language model: V4 Flash and V4 Pro. Both use a mixture-of-experts architecture, activating only a subset of parameters per task to reduce inference costs. The Pro model boasts 1.6 trillion total parameters (49 billion active), surpassing Moonshot AI's Kimi K 2.6 and MiniMax's M1, making it the largest open-weight model available. The Flash variant has 284 billion parameters (13 billion active). Each model supports a 1-million-token context window, enough for large codebases or documents.

DeepSeek claims V4 Pro matches GPT-5.4 on coding competition benchmarks and outperforms OpenAI's GPT-5.2 and Gemini 3.0 Pro on some reasoning tasks. However, it lags behind frontier models on knowledge tests, a gap the lab estimates at 3–6 months. Both models are text-only, unlike multimodal peers. Pricing is aggressive: Flash costs $0.14/M input tokens and $0.28/M output, undercutting GPT-5.4 Nano and Claude Haiku 4.5. Pro costs $0.145/M input and $3.48/M output, beating Gemini 3.1 Pro and GPT-5.5. The launch follows U.S. accusations of IP theft by Chinese labs, with Anthropic and OpenAI previously accusing DeepSeek of model distillation.