DeepSeek has released V4, a one-trillion-parameter Mixture-of-Experts model with fully open weights, meaning anyone can download and run it. The model achieves benchmark scores competitive with the top US frontier models — including GPT-5.4 and Claude Opus 4.7 — while costing an estimated $5.2 million to train, a figure that is orders of magnitude smaller than what US labs typically spend. DeepSeek achieved this efficiency by activating only a subset of its parameters for any given task, a technique that lets the model punch well above its computational weight.
The release is significant on two levels. First, it demonstrates that open-weight models are no longer meaningfully behind proprietary ones on capability benchmarks. Second, it raises pointed questions for policymakers: if a Chinese lab can match frontier performance for a fraction of the cost and then publish the weights freely, US export controls on chips alone cannot contain the spread of cutting-edge AI. The model is already being downloaded and fine-tuned by researchers worldwide.
The timing is notable given that, just weeks ago, OpenAI, Anthropic, and Google jointly accused DeepSeek of using fraudulent accounts to scrape training data from their proprietary models — a practice called adversarial distillation. DeepSeek has not publicly responded to those accusations. Whether the V4 weights were trained from scratch or partly distilled from closed models remains an open question in the research community.
For students learning about AI, DeepSeek V4 is a useful reminder that capability is not the exclusive domain of the richest labs. The open-weight release means you can experiment with a frontier-tier model on your own hardware or a cheap cloud instance. More broadly, it shows that understanding how AI systems are built — not just how to use them — is an increasingly valuable skill, because the underlying techniques are becoming more accessible and more consequential every year.