Chinese AI lab DeepSeek has announced V3.2, the latest addition to its V3 model series. Launched on Monday, this new model comes in two versions: “Thinking” and the more powerful “Speciale.” DeepSeek V3.2 aims to further advance the capabilities of open-source AI while offering a significantly lower cost compared to its competitors. Accessible through Hugging Face, the model aims to keep accessible AI competitive for developers.
New balances in the world of AI: What does DeepSeek V3.2 offer?
According to data published by the company, V3.2 Speciale outperforms industry-leading closed-source models like OpenAI’s GPT-5 High, Anthropic’s Claude 4.5 Sonnet, and Google’s Gemini 3.0 Pro in some reasoning tests. Despite this performance increase, the cost difference is quite striking. For example, while Gemini 3 API access can reach up to $4 per million tokens, this fee is only $0.028 for V3.2 Speciale. The model is also reported to have achieved gold-level performance at the International Mathematical Olympiad and the International Informatics Olympiad.
DeepSeek has made significant changes to its technical infrastructure to address the performance issues of open-source models and reduce processing overhead. To address the slowness of the traditional “vanilla attention” mechanism, a new mechanism called DeepSeek Sparse Attention (DSA) has been developed. This system first performs a quick scan of the data and then focuses on relevant sections. This can be likened to searching through an organized library by going directly to the relevant section, rather than sifting through a scattered stack of books to find a sentence.
Despite these advantages offered by the model, the researchers highlight some limitations. According to the published article, the new model’s “world knowledge,” or general knowledge, is more limited than leading closed-source models. Furthermore, it can sometimes struggle with complex tasks and consume more tokens to produce output of comparable quality to its competitors. Nevertheless, the ability of an open-source model developed by a China-based company to compete with the world’s most advanced closed-source models is changing the perception of the “performance gap” in the industry.
This development calls into question the necessity of high-cost closed-source systems and symbolizes the rise of powerful alternatives that developers can access virtually free of charge. If open-source models continue to develop at this rapid pace and close the gap, user motivation to pay high prices for closed-source systems could significantly diminish. Do you think open-source, low-cost AI models will completely replace the paid tech giants in the future?
{{user}} {{datetime}}
{{text}}