This video discusses the updated DeepSeek R1 model, called Deepseek R10528, which was released on May 28, 2025.
The key points mentioned in the video are:
-
Model Update & Performance: The R10528 is a minor trial upgrade that significantly enhances code generation capabilities. In programming benchmark tests, it performs similarly to OpenAI’s GPT-4 mini and GPT-3 models, outperforming XAI’s Grok-3 Mini and Alibaba’s Qianwen 3.
-
Licensing & Availability: The model is fully open-source under the MIT license, allowing for commercial use and the creation of derivative models. It is accessible through various platforms, including the DeepSeek website, Hugging Face, AWS via Amazon Bedrock and SageMaker, and Nibbius AI Studio.
-
API Pricing: The API pricing is highly competitive at around 14 cents per million input tokens, which is significantly lower than OpenAI’s $7.50 per million tokens.
-
DeepSeek R2 and Future Strategy: The much-anticipated DeepSeek R2 has not yet been released. The speaker suggests that the delay is likely due to the team aiming for a major breakthrough and the challenges of adapting to a new ecosystem with domestic AI chips like Huawei’s, especially given the US AI product ban which limits access to advanced computing clusters. The speaker believes DeepSeek is pursuing a dual-track approach, developing both a domestic and an international ecosystem to maintain its market position.
-
DeepSeek’s Market Position: DeepSeek has become a “de facto standard” in China, with its models adopted by vehicle AI systems, government agencies, financial institutions, and other industries. The speaker suggests that DeepSeek’s focus should be on cost-effectiveness, openness, and compatibility rather than blindly pursuing absolute performance leadership.