DeepSeek, the China-based AI firm, has launched two new reasoning-capable AI fashions, V3.2 and V3.2-Speciale, with the latter outperforming Gemini 3.0 Professional and GPT-5 Excessive in benchmarks, the corporate stated in information launch on Monday. In the meanwhile, V3.2 is offered on the app and net, whereas V3.2-Speciale can solely be accessed by way of the API.
DeepSeek V3.2 is a follow-up to V3.2-Exp, which stands for experimental and was launched in September. The Hangzhou-based AI firm says V3.2 surpasses the efficiency of OpenAI’s GPT-5 in reasoning benchmarks and has reasoning efficiency “on par” with Gemini 3.0 Pro. DeepSeek did not launch testing towards GPT-5 Professional. That could possibly be as a result of ChatGPT is blocked in China. What’s extra, DeepSeek stated it is dedicated to staying open supply, that means any firm can load DeepSeek’s fashions totally free. However contemplating V3.2 has 685 billion parameters, it might solely run on a large server stack with hundreds of thousands of {dollars} in {hardware}.
🏆 World-Main Reasoning
🔹 V3.2: Balanced inference vs. size. Your each day driver at GPT-5 degree efficiency.
🔹 V3.2-Speciale: Maxed-out reasoning capabilities. Rivals Gemini-3.0-Professional.
🥇 Gold-Medal Efficiency: V3.2-Speciale attains gold-level leads to IMO, CMO, ICPC World… pic.twitter.com/ZFGzDMshBy— DeepSeek (@deepseek_ai) December 1, 2025
Do not miss any of our unbiased tech content material and lab-based critiques. Add CNET as a most well-liked Google supply.
DeepSeek did not instantly reply to a request for remark.
DeepSeek’s launch of three.2 comes as the corporate continues to place American AI firms on discover. The corporate first gained worldwide consideration earlier this year when it launched a reasoning-capable mannequin on-line, totally free.
It instantly turned the narrative about reasoning fashions on its head, exhibiting that it is attainable to run smarter, extra succesful AI fashions at a fraction of the associated fee. This spooked Wall Road, with some questioning if OpenAI, Google and Anthropic weren’t innovating quick sufficient.
Since then, ChatGPT, Gemini and Claude have all launched reasoning-level fashions totally free, with higher-level “pondering” fashions being accessible for paid subscribers.
DeepSeek-V3.2-Speciale gained gold within the 2025 Worldwide Mathematical Olympiad (IMO) and Worldwide Olympiad in Informatics (IOI), the corporate stated.
(Disclosure: Ziff Davis, CNET’s father or mother firm, in April filed a lawsuit towards OpenAI, alleging it infringed Ziff Davis copyrights in coaching and working its AI techniques.)
