DeepSek Makes Silicon Valley Panic: V3.2 And V3.2-Special Claims More Than GPT-5!

Liang Wenfeng, founder of China's artificial intelligence startup (AI) DeepSek (photo: x @AskPerplexity)

JAKARTA - DeepSeek has hit the artificial intelligence industry again in a way that feels like a surprise sequel: a year after its model created a global fast market, the Chinese company released two new models DeepSek V3.2 and V3.2-Speciale. This version is claimed to be able to challenge and even overtake the current most powerful AI system, including OpenAI's GPT-5 and Gemini 3 Pro from Google.

The strategy is still the same: not just chasing the size of the model, but efficiency. When US labs rely on the most up-to-date and supermajor chip clusters, DeepSeek actually insists that a more efficient training approach can bring equivalent intelligence without the Sultan's class infrastructure.

The company also emphasizes that the standard version of the model has been equipped with the ability to reason for the use of tools natively, so users do not need to activate a special reasoning mode.

The biggest highlight fell on the V3.2-Speciale, a model that DeepSek says has surpassed GPT-5 in an internal benchmark and fought the equivalent of the Gemini 3 Pro in reasoning-based tough tasks.

As evidence, DeepSeek publishes the results of its model performance at the International Mathematical Olympiad 2025 and International Olympiad in Informatics, complete with final entry that can be examined by the public.

The performance leap is associated with two major innovations: a spread attention mechanism designed for long context efficiency, as well as an expanded and trained pipeline enforcement learning using more than 85,000 multi-step complex tasks. All of these tasks are generated through their in-house system called the agency task synthesis.

DeepSek V3.2 is already available to the public through the website, mobile app, and API. However, the V3.2-Special has just opened via the API endpoint, which will be closed after December 15, 2025. This model is currently running as a reasoning-only machine without tool calling capabilities.

Industry is still waiting for independent benchmarks to confirm this big claim. But one thing is getting clearer: Liang Wenfeng's first AI wants to prove that the quality of the top AI doesn't have to come at a short breathing price. That message alone is enough to make big industry players start reviewing the way they fight in an increasingly competitive field of technology.