DeepSeek V3.2. What’s New in Chinese AI Models? – AI News – #2 December 2025

The Chinese AI scene has just made a move that could shake the dominance of Western tech giants. The release of the DeepSeek V3.2 models and its specialized V3.2-Speciale version is a clear signal that the race to create the most advanced reasoning model is far from slowing down. These new models not only challenge GPT-5 and Gemini 3.0 Pro but do so in an open-source format and at significantly lower operational costs.

DeepSeek has brought two variants of its latest achievement to market. The base model, DeepSeek V3.2, is designed as a high-efficiency solution balancing performance and costs. Meanwhile, the DeepSeek V3.2-Speciale variant is a powerhouse tuned for maximizing reasoning capabilities, which in tests matches or even surpasses flagship models from Google and OpenAI.

According to the technical report, the standard DeepSeek V3.2 achieves results comparable to the GPT-5-High model in reasoning-intensive tasks. The Speciale variant performs even more impressively. Thanks to advanced reinforcement learning (RL) protocols and an increased computational budget in the post-training phase, this model reaches a level close to Gemini-3.0-Pro.

It’s worth noting that the Speciale version earned “gold medals” in prestigious competitions such as the International Mathematical Olympiad (IMO) 2025 and the International Olympiad in Informatics (IOI).

DeepSeek didn’t stop at just boosting computational power. Significant changes to the architecture were introduced to address key issues in today’s LLMs.

One of the main breakthroughs is the introduction of the DeepSeek Sparse Attention (DSA) mechanism. Traditional attention mechanisms become inefficient with very long contexts. DSA drastically reduces computational complexity, allowing the model to efficiently process long sequences without performance loss. This is a key change for business and analytical applications, where the “context window” is often a bottleneck.

The most intriguing novelty from the perspective of building autonomous AI agents is the “Thinking in Tool-Use” concept (thinking during tool use). Previous models often lost their reasoning trace when invoking an external tool (e.g., a code interpreter or search engine), forcing them to reprocess the entire problem.

DeepSeek V3.2 introduces a context management system that:

DeepSeek’s strategy of releasing models as open source (under the MIT license) poses a direct challenge to the closed ecosystems of OpenAI or Anthropic. The company has released the model weights on the Hugging Face platform, allowing developers to deploy them freely.

The new models are not only efficient but also cheaper to operate—the Speciale version API is priced significantly below competitors’ rates, and the standard V3.2 is intended to be a “daily driver” for users. This also demonstrates that, despite sanctions on advanced integrated circuits (GPUs), the Chinese AI sector is finding ways to optimize architecture and scale computing power. However, regulatory issues must be kept in mind—in the past, DeepSeek models faced barriers in Europe due to data privacy concerns.

The premiere of DeepSeek V3.2 is proof that the gap between open source models and closed “frontier” systems is not only not widening, but in some areas is beginning to vanish. For the SEO and marketing industry, this means cheaper access to tools that generate content and code at the highest global level.

Want to know what happens next? Stay up to date with news from the world of AI and subscribe to the Delante newsletter!

Source of information about DeepSeek V3.2: https://api-docs.deepseek.com/news/news251201