DeepSeek launches two new AI models: DeepSeek-V3.2 functions as a daily reasoning assistant, and DeepSeek-V3.2 Special focuses on heavier reasoning tasks.
The Chinese AI startup DeepSeek is coming up with two new AI models. These might just dare to get in the way of OpenAI’s GPT-5. At least, that is what the Chinese company itself indicates. DeepSeek-V3.2 replaces the earlier V3.2-Exp version and is available via app, web and API. The Special variant focuses on heavier reasoning tasks and can only be used via a temporary API for the time being.
AI reasoning
According to DeepSeek, V3.2 delivers a balance between reasoning capacity, speed and context length. The model should position itself in the same category as other recent leading generative models.
V3.2-Special can process more tokens and perform longer thought steps, but therefore consumes more computing power per request. This variant does not support tool-calls and will remain available until mid-December 2025 for evaluation
Agents
The models have been developed with a view to use in agent environments. DeepSeek describes its own method for data synthesis, based on more than 1,800 environments and approximately 85,000 complex instructions. With this, the company wants to train agents that can plan multiple steps, deploy tools and complete tasks autonomously.
read also
Latest DeepSeek Model Shows Future without CUDA
A striking element is the integration of ‘thinking mode’ in tool-use. According to the documentation, V3.2 can call tools both in thinking mode and in a more direct mode. That should give developers more control over how an agent reasons, stores intermediate steps and substantiates decisions.
The technical report, available via Hugging Face, describes the architecture, training strategy and benchmark results. DeepSeek claims that V3.2-Special achieves gold-level results in various international programming and mathematical competitions, such as IMO-, ICPC- and IOI-like benchmarks.
Open source available
In addition to the commercial API, DeepSeek also makes both models available as open source. DeepSeek-V3.2 and DeepSeek-V3.2-Special can be downloaded via Hugging Face for use in your own infrastructure.
For companies and developers in the Benelux, this opens the door to on-premises implementations, experimental agent workflows and customization on top of the standard models.
Not for federal government
Although the new models sound promising, not everyone will be able to enjoy them. As of December 1, staff of the federal government will no longer be allowed to use applications from the Chinese DeepSeek. All applications from the AI company must therefore be removed from staff members’ devices. Questions have been raised for some time regarding the privacy of the model, given its Chinese roots.
read also