Tech Desk 
Dec 04, 2025
Chinese startup DeepSeek has intensified the global AI race by releasing two new large language models, DeepSeek-V3.2 and DeepSeek-V3.2-Special.
Both of the models are fully free, open-source rivals to GPT-5.1 and Gemini 3 Pro.
DeepSeek-V3.2 and DeepSeek-V3.2-Special model
The V3.2 model targets everyday reasoning tasks, while the Special variant focuses on advanced math and coding.
Both models are available under the MIT open-source license, allowing anyone to run, modify, or commercialise them.
The open approach contrasts with how major US firms such as OpenAI and Google have made access only through paid Application Programming Interfaces (APIs) and restricted weight releases for safety reasons.
DeepSeek claims that the new systems match or outperform leading Western models on long-form reasoning, tool use, and competitive benchmarks.
The Special model achieved stand-out scores: 99.2% at the Harvard-MIT Math Tournament, gold-level in coding competitions, and strong bug-fixing accuracy, all without using external tools or internet access.
One big advance is the company's DeepSeek Sparse Attention (DSA) architecture. While traditional transformers scale their costs with growing context length, DSA keeps its focus only on the most relevant parts of a document, thereby reducing long-context compute expenses by as much as 70%.
Models with a 128,000-token context window and no licensing restrictions lower the barrier for developers and small teams who have previously needed large cloud budgets to work with systems of this grade.
DeepSeek has also enhanced tool-use reasoning. In simple terms, the models remember all previously used tools instead of being reset like most AI agents do after every step.
Trained on over 85,000 complex synthetic tasks involving browsers and coding environments, the new systems are capable of doing multi-stage tasks such as budget trip planning, rate exchange checking, and running code themselves.
But the company's fully open release is drawing regulatory scrutiny. Germany recently moved to block the app over data-transfer concerns.
Italy banned it earlier this year, and US lawmakers are pushing to restrict its use on government devices, citing security risks linked to its Chinese origin.
