DeepSeek V4: The Chinese AI Challenges American Giants

DeepSeek Unveils Its New V4 Models

The Chinese artificial intelligence lab DeepSeek has recently launched two new versions of its large language model, the DeepSeek V4. These models, named V4 Flash and V4 Pro, succeed last year's V3.2 model and the reasoning model R1, which had already made waves in the AI sector.

The V4 models from DeepSeek utilize a technique called mixture of experts, allowing only certain parameters to be activated for each task, thereby reducing inference costs. Each model features a context window of 1 million tokens, making it easier to integrate large codebases or documents into prompts.

Impressive Performance and Size

The V4 Pro model stands out with its 1.6 trillion parameters, of which 49 billion are activated, placing it at the forefront of open-weight models. It surpasses the Kimi K 2.6 from Moonshot AI and the M1 from MiniMax. The smaller V4 Flash model has 284 billion parameters with 13 billion active.

DeepSeek claims that these models are more performant and efficient than their predecessor, the V3.2, thanks to architectural improvements. They are said to have nearly closed the gap with current leading models, whether open or closed, particularly on reasoning benchmarks.

Comparison with Leading Models

DeepSeek's V4-Pro-Max model reportedly outperforms its open-source competitors on certain reasoning benchmarks and competes with GPT-5.2 and Gemini 3.0 Pro. In coding competitions, the performance of the V4 models is said to be comparable to that of GPT-5.4.

However, the V4 models show a slight lag in knowledge tests compared to leading models like GPT-5.4 from OpenAI and Gemini 3.1 Pro from Google. This lag is estimated to be around 3 to 6 months behind market leaders.

Limitations and Cost

The V4 Flash and V4 Pro models only support text, unlike many closed-source counterparts that offer support for audio, video, and image understanding and generation. One of the major advantages of the V4 models is their cost. The V4 Flash model is priced at $0.14 per million input tokens and $0.28 per million output tokens, while the V4 Pro is at $0.145 per million input tokens and $3.48 per million output tokens. These rates are lower than those of many leading models, such as GPT-5.4 Nano, Gemini 3.1 Flash, and Claude Haiku 4.5.

Context of International Tensions

The launch of these models comes amid heightened tensions, just a day after the United States accused China of large-scale intellectual property theft in the field of AI. DeepSeek has been specifically accused by Anthropic and OpenAI of "distillation," meaning copying their AI models.