DeepSeek’s newest AI models V3.2 and V3.2-Speciale have arrived, and they’re gunning straight for the likes of GPT-5 and Gemini 3.0 Pro.

DeepSeek V3.2 models go beyond expectations

In January, DeepSeek shocked global markets with an earlier release. Now, the company is back with two fresh contenders. The regular V3.2 serves as a solid workhorse: fast, adaptable, and built for daily use. But don’t let the modest label fool you, it introduces tool-use features that blend in the model’s internal reasoning without slowing things down.

The Speciale version, however, isn’t playing it safe. It’s engineered to outthink even the flashiest frontier models. DeepSeek says V3.2-Speciale beats GPT-5 in several core reasoning tasks and goes toe-to-toe with Gemini 3.0 Pro.

DeepSeek V3.2-Speciale takes on pure reasoning

Where most companies throw compute at the problem, DeepSeek has been more surgical. Speciale was trained using a custom sparse attention method called DeepSeek Sparse Attention (DSA), which cuts down the weight of long-context processing. On top of that, the company built a reinforced learning system that scales far more efficiently than its peers.

If you’re wondering whether it’s all smoke, consider this: Speciale hit top marks at both the International Mathematical Olympiad and the Informatics Olympiad. DeepSeek even published its competition submissions, so the public can see the math for themselves.

Agentic thinking shifts the game

Instead of just mimicking commands, V3.2 models aim to think through them. DeepSeek trained these models using a Large-Scale Agentic Task Synthesis Pipeline, feeding in over 85,000 compound tasks. This setup allows the models to handle multi-step instructions without needing constant prompting or fine-tuning.

That agentic twist could prove powerful for real-world deployment. Whether it’s technical workflows or open-ended queries, the model is built to adapt in motion.

What makes DeepSeek different

Here’s where DeepSeek stands out:

• Prioritizes training efficiency over raw compute
• Integrates tool use within both thinking and non-thinking tasks
• Leverages sparse attention for long-context gains
• Released competition results for external audit
• Trained on a staggering number of synthetic agent tasks

Most labs build models for benchmarks. DeepSeek’s are being tested in the wild.

DeepSeek V3.2 availability details

V3.2 is now available on mobile, web, and through API access. The Speciale version, however, is API-only and comes with a countdown it’s set to vanish after December 15, 2025. One catch: it won’t support tool use, making it a raw reasoning engine by design.

Agentic thinking meets real competition

DeepSeek’s approach is making the giants sweat. By threading intelligence into practical AI workflows, it’s proving that high performance doesn’t have to come with sky-high compute bills. The AI wars aren’t over but the quiet ones might just be winning.