DeepSeek-V3.2 Unveiled: A Strong Challenger to GPT-5
DeepSeek unveils V3.2 and its Speciale edition, boosting everyday performance and reasoning. Speciale outperforms rivals like Gemini 3.0 Pro and GPT-5 on key benchmarks, with API access.
DeepSeek, a Chinese AI start-up, has unveiled two language models: DeepSeek-V3.2 and its enhanced variant DeepSeek-V3.2 Speciale. The standard model builds on the earlier V3.2-Exp release and aims to be a reliable tool for everyday work. The developers say it matches the speed and quality of answers seen from the GPT-5 model.
Speciale edition sharpens reasoning
Speciale emphasizes stronger logical thinking and problem solving. In several benchmark tests, it beat Gemini 3.0 Pro and GPT-5 High. Notable gains occurred in math and science challenges run by AIME and HMMT, as well as the IMO Answer Bench with hundreds of contest problems.
Beyond math, Speciale showed robust performance in GPQA Diamond, covering physics, biology, and chemistry, and in the Last Human Exam (HLE) tasks that mix questions from exact sciences and humanities. The company notes that these enhanced capabilities demand substantial computing power and token usage, so access to Speciale is currently API-only.
Availability and access
DeepSeek-V3.2 is already available on the web, in mobile apps, and via API. Developers can view the code for both models on the Hugging Face platform.
Expert perspective
Expert opinion: Industry researchers say Speciale marks a meaningful jump in reasoning ability, though practical deployment depends on infrastructure and cost management. In short, higher performance comes with greater resource needs.
Summary
DeepSeek has released two members of its V3.2 family. The standard model provides reliable daily performance for a wide range of tasks, while Speciale targets advanced reasoning and has outperformed several leading rivals on benchmark tests. Access to Speciale is API-based due to its resource demands, while the standard version remains available on the web, in apps, and via API. Both models can be explored through Hugging Face.
Key insight: Advanced reasoning breakthroughs like Speciale show the potential for AI to surpass top models in tests, but real-world use hinges on scalable API access and robust infrastructure.