Artificial intelligence startup DeepSeek has launched its most powerful model variant, DeepSeek-V3.2-Speciale, which it says can match Google DeepMind’s new Gemini 3 Pro model in some tasks, despite the Chinese company’s limited access to advanced semiconductor chips.
The open source lab’s achievement sparked widespread discussion in the artificial intelligence research community because it coincides with the prestigious annual conference on Neural Information Processing Systems (NeurIPS).
Hangzhou-based DeepSeek said in an announcement on Monday that V3.2-Speciale is comparable in inference capabilities to the Gemini 3 Pro released by Google two weeks ago, while the performance of V3.2-Speciale’s base model (V3.2 launched on the same day) is comparable to OpenAI’s GPT-5 launched in August.
Do you have questions about the hottest topics and trends from around the world? Get the answers with SCMP Knowledge, a new platform of curated content with explainers, FAQs, analyzes and infographics brought to you by our award-winning team.
DeepSeek said that V3.2-Speciale achieved gold medal performance in the International Mathematical Olympiad test, a result that had previously been achieved only by internal models of OpenAI and Google DeepMind, but has not yet been made public.
DeepSeek-V3.2-Speciale is the startup’s most powerful AI model variant to date. Photo: Handout alt=DeepSeek-V3.2-Speciale is the startup’s most powerful AI model variant to date. Photo: Handout >
DeepSeek has open sourced the V3.2 model on the developer platform Hugging Face. However, V3.2-Speciale will only be accessible through an application programming interface, or API, due to its “higher token usage,” according to the announcement.
On social media, Google DeepMind principal research engineer Susan Zhang praised DeepSeek for providing a detailed technical report on its new model, noting the startup’s efforts in stabilizing the model after training and enhancing its agent capabilities.
DeepSeek achieved this despite having fewer “total training failures” than its U.S. rival, according to its technical report. FLOPs (floating point operations per second) are a standard measure of computing power used to train artificial intelligence models.
DeepSeek acknowledged that V3.2 still lags behind leading US models such as Gemini 3 Pro in terms of token efficiency and world knowledge, but said it intends to close these gaps by expanding the computing power used for pre-trained models.
These new models, launched ahead of NeurIPS, were compared to OpenAI’s release of ChatGPT in late November 2022, which also coincided with that year’s conference in New Orleans.
DeepSeek, which has always been low-key, did not disclose whether it would send employees to participate in this year’s conference. Florian Brand, an expert on China’s open source artificial intelligence ecosystem who attended this year’s NeurIPS in San Diego, said any DeepSeek researchers present would likely draw significant attention.
“After DeepSeek announced this news, all the group chats were full today,” he said.
This year’s conference was held for the first time in two locations, San Diego and Mexico City, due to concerns about international researchers obtaining U.S. visas. Many Chinese contestants chose Mexico City as the venue.
This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice on China and Asia for more than a century. For more SCMP stories, explore the SCMP app or visit SCMP on Facebook and twitter Page. Copyright © 2025 South China Morning Post Publishing Limited. All rights reserved.
Copyright (c) 2025. All rights reserved by South China Morning Post Publishing Limited.
