China’s DeepSeek AI just dropped a bombshell in the tech world. Brace yourselves, folks. The company has revealed its V3.2-Speciale model, claiming it rivals Google’s Gemini 3 Pro in reasoning capabilities. Yes, you heard that right. In what feels like a scene straight out of an AI movie, DeepSeek is shaking the cobwebs off the industry. The announcement came during the NeurIPS conference, where the air buzzed with excitement and disbelief.
DeepSeek AI shakes the tech world with its V3.2-Speciale, claiming to rival Google’s Gemini 3 Pro!
But wait, there’s more. This V3.2 model doesn’t just match the big players; it also performs on par with OpenAI’s GPT-5, launched just last August. Talk about impressive timing. V3.2-Speciale achieved a gold-medal performance on the International Mathematical Olympiad test, a feat only previously reached by the crème de la crème—OpenAI and Google DeepMind’s internal models. Moreover, the launch coinciding with the NeurIPS conference has sparked discussions within the AI research community.
What’s the catch? The Speciale variant is only available through API due to its insane resource demands. Meanwhile, the base model V3.2 is open-sourced on Hugging Face. So, if you’re a coder with dreams of AI glory, you might just hit the jackpot. But let’s be real, it took around $6 million to develop this masterpiece compared to the whopping $100 million spent on GPT-4. Efficiency at its finest!
DeepSeek’s performance claims are bold. The V3.2-Speciale equals Gemini 3 Pro in reasoning tasks despite limited chip access. It also has 63% fewer mistakes than general AI models. How? They focused heavily on industry-specific training. And guess what? It’s bilingual in English and Chinese, trained on a staggering 8.1 trillion tokens. Additionally, this new model’s engineering efficiency is noted as impressive in the competitive landscape.
But hold your applause. Alphabet shares took a hit, falling 1.65% on announcement day. Ouch. DeepSeek is challenging the US-dominated AI landscape with its low-cost, open-source models. This raises eyebrows about export controls and the future of AI infrastructure.
Critics aren’t silent either. Sure, they’ve got some impressive stats, but there are notable gaps in token efficiency and general knowledge compared to US rivals. Plus, compliance with Chinese regulations means they’re censoring politically sensitive topics. Not exactly the open frontier some might hope for.
