Chinese AI startup DeepSeek has released an updated version of its flagship R1 reasoning model on Hugging Face, marking a continued push toward open-source development even as scrutiny over its capabilities intensifies abroad. The release was announced via WeChat on Wednesday, with the model now accessible to developers under the permissive MIT license, allowing for full commercial use.
This new iteration is described by DeepSeek as a “minor” upgrade, but the technical specifications are far from modest. The updated R1 weighs in at 685 billion parameters—placing it among the largest publicly released language models to date. While the Hugging Face repository offers little in terms of documentation, it includes the necessary configuration files and raw model weights, giving developers the tools to implement or fine-tune the model for their own use cases.
At its current size, R1 is not designed for consumer hardware. Running or fine-tuning a model of this scale requires high-end computing infrastructure, typically reserved for research labs, enterprise environments, or well-funded AI ventures. However, its open licensing and availability signal a notable step toward transparency, especially at a time when many leading AI models remain gated behind closed APIs or proprietary platforms.
DeepSeek first made headlines earlier in 2025 when it launched the original version of R1, a high-performance model that demonstrated strong reasoning and multi-step task-solving abilities—areas that remain challenging for most large language models. The release drew attention not only for its technical performance but also for the company’s rapid rise in a landscape dominated by U.S.-based firms like OpenAI, Anthropic, and Google DeepMind.
The R1 series has also attracted the attention of U.S. officials, some of whom have raised concerns about its potential national security implications. With geopolitical tensions influencing global AI policy, open access to high-capacity reasoning models developed outside of Western jurisdictions has become a point of contention. Critics argue that such models could be adapted for surveillance, misinformation, or strategic applications, while proponents maintain that open science and competition remain essential for progress.
The release of the R1 model under an MIT license could signal DeepSeek’s intent to foster international collaboration, or at least to position itself as a legitimate player in the broader AI ecosystem. With tools like Hugging Face making distribution and experimentation easier than ever, DeepSeek may be seeking to build developer trust through accessibility—even if the model’s scale means only a limited number of organizations will be able to deploy it effectively.
While this version of R1 is labeled a minor update, its availability opens new doors for researchers and developers exploring cutting-edge reasoning capabilities. As AI companies continue to weigh the trade-offs between openness and control, DeepSeek’s decision to release its most advanced work to date under such a liberal license is likely to keep it in the global spotlight—for both technical and political reasons.
