So DeepSeek woke up and chose violence — dropping not one, but two new AI models with benchmarks so spicy, they’re basically subtweets aimed at Google and OpenAI.
Here’s the rundown:
They launched V3.2 and its bigger sibling, V3.2-Speciale, and like I said earlier, the numbers are wild.
For context: V3.2 is the follow-up to that experimental model they shipped back in September — but this new one? DeepSeek claims it outperforms GPT-5 High on several reasoning benchmarks.
And Speciale? That version straight-up outperforms GPT-5, goes toe-to-toe with Google’s Gemini-3 Pro, and hits gold-medal-level scores in competitions like the International Math Olympiad and the International Olympiad in Informatics.
Again — their words, not mine.
But here’s where it gets extra interesting — DeepSeek isn’t just pushing raw intelligence. They’re going all-in on tool integration, which means the model doesn't just “think”, it can think while using calculators, search engines, and code executors.
DeepSeek even says this is their first model where thinking is directly wired into tool use, and it can switch between thinking and non-thinking modes.
Translation? It can be deep and analytical… or fast and practical.
Then there’s Speciale, which is basically the graduate student version — built for long-form reasoning and advanced math. It handles multi-step logic without collapsing halfway through.
If V3.2 is the everyday genius, Speciale is the one you call when your research paper starts hallucinating.
Oh — and they’re keeping it open-source.
Which sounds awesome… until you remember the model has 685 billion parameters, meaning you’re only running this thing if you own a warehouse, a power plant, and maybe a small moon.
And just to stack the announcement even higher, DeepSeek also revealed a brand-new method for training AI agents — the kind that can act independently, make decisions, and operate without hand-holding.
Basically, they’re prepping for the era of “AI employees” that don’t need a manager… which, depending on who you are, is either super exciting or mildly terrifying.
So what does all of this mean?
It means DeepSeek isn’t just catching up, they’re firing shots at the absolute top of the AI leaderboard.
So yeah, the next phase of the AI race might not be just OpenAI vs Google… it might be OpenAI vs Google vs DeepSeek.
And honestly? We’re here for it!
Right now, V3.2 is available on both the app and web, while V3.2-Speciale is API-only for now.
Want the full details? Check it out here.
