Key facts
- DeepSeek open-sourced its DeepInfer framework for efficient AI model inference.
- The framework supports DeepSeek's V4.1-Flash model on Nvidia Blackwell GPUs and Huawei hardware.
- DeepSeek V4 is priced significantly lower than comparable US models.
- US officials accuse Chinese AI firms of using illicit distillation attacks to replicate US model capabilities.
Chinese artificial intelligence startup DeepSeek has fully open-sourced its DeepInfer framework, designed to run AI models efficiently. The company's senior researcher described the growing divide between open-source and closed-source AI as a battle between "light and darkness." The framework supports DeepSeek's V4.1-Flash model on Nvidia's SM100 Blackwell GPUs and Huawei hardware. US government officials have asserted that V4 was trained on smuggled Nvidia Blackwell chips, which are banned in China. They also accuse Chinese AI firms, including DeepSeek, of using illicit distillation attacks against US models. DeepSeek admits it cannot serve its V4 Pro model to most customers due to a lack of chips, resulting in a model that is worse than leading US models and more expensive than its Chinese competitors, though it offers a one-million-token context window at a fraction of V3's inference compute cost. Despite performance gaps, V4 is open source, large-scale, and priced for mass deployment, potentially offering significant competitive value.