What’s the story?
Chinese artificial intelligence (AI) startup DeepSeek has launched two new models: DeepSeek-V3.2 and DeepSeek-V3.2-Speciale. The move comes as a direct challenge to the dominance of Google’s Gemini series and OpenAI’s ChatGPT in the AI chatbot space. The company previously made a name for itself with its low-cost DeepSeek R1 and V3 models, which rival ChatGPT’s performance.
DeepSeek-V3.2: Inference-first model
According to the company, the new DeepSeek-V3.2 is an “inference-first model built for agents.” This model claims to provide performance comparable to GPT-5 for common tasks. Another variant, DeepSeek-V3.2-Speciale, is said to have inference capabilities comparable to Google’s latest model Gemini 3 Pro, especially in complex problem-solving scenarios.
DeepSeek-V3.2-Speciale performance at international competitions
The DeepSeek-V3.2-Speciale model has also shown excellent performance in international competitions such as the International Mathematics Olympiad (IMO) and the International Olympiad in Information (IOI). This achieved gold medal-level performance, a feat previously achieved only by proprietary models from Google and OpenAI. This model outperformed both GPT-5 High and Gemini 3 Pro on pure math benchmarks such as AIME and HMMT, and also on pure coding logic tasks.
Performance of DeepSeek-V3.2-Speciale in humanity’s final test
In Humanity’s Last Exam (HLE), one of the most demanding benchmarks, DeepSeek-V3.2-Speciale outperformed GPT-5 High but fell short of Gemini 3 Pro. The new model is now available on the DeepSeek website and app. However, please note that DeepSeek V3.2-Speciale is currently only available via API, and there is no announcement as to when it will be accessible to general users.
