Chinese AI lab DeepSeek has released two preview versions of its latest large language model, DeepSeek V4, an eagerly anticipated upgrade from last year’s V3.2 model and its R1 reasoning model that captivated the AI industry.
The company reports that both DeepSeek V4 Flash and V4 Pro are mixture-of-experts models with context windows of 1 million tokens each, sufficient for incorporating extensive codebases or documents in prompts. This approach activates only certain parameters per task to reduce inference costs.
The Pro model features 1.6 trillion parameters (49 billion active), making it the largest open-weight model, surpassing Moonshot AI’s Kimi K 2.6, MiniMax’s M1, and more than doubling DeepSeek V3.2. The smaller V4 Flash has 284 billion parameters (13 billion active).
DeepSeek asserts these models are more efficient and effective than V3.2 due to architectural advancements and have nearly matched leading models, both open and closed, in reasoning benchmarks.
The company claims its V4-Pro-Max model surpasses its open-source counterparts in reasoning benchmarks and exceeds OpenAI’s GPT-5.2 and Gemini 3.0 Pro in some tasks. In coding competitions, DeepSeek states both V4 models’ performance is comparable to GPT-5.4.
However, the models slightly lag behind frontier models in knowledge tests, specifically against OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro, indicating a 3 to 6-month developmental delay behind cutting-edge models.
Both V4 Flash and V4 Pro support text only, unlike many closed-source competitors that offer audio, video, and image capabilities.
Significantly, DeepSeek V4 is more affordable than current frontier models. The V4 Flash model costs $0.14 per million input tokens and $0.28 per million output tokens, undercutting several competitors. The V4 Pro model costs $0.145 per million input tokens and $3.48 per million output tokens, also more cost-effective than rivals.
The launch coincides with accusations from the U.S. of China stealing American AI labs’ IP on a large scale using proxy accounts. DeepSeek has been accused by Anthropic and OpenAI of “distilling” their AI models.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.