DeepSeek unveils V3.2 and V3.2-Speciale to rival GPT-5 and Gemini 3 Pro

DeepSeek has once again seized the spotlight, unveiling two new models — V3.2 and V3.2-Speciale. A year after the previous version made waves, the company returns with ambitious claims: by its internal evaluations, the newcomers can compete with — and in some cases edge past — flagship AI systems like OpenAI’s GPT-5 and Google’s Gemini 3 Pro. True to its “efficiency over scale” mantra, DeepSeek maintains that its training methods deliver strong intellectual performance on far less expensive hardware. Even the base model, the company notes, comes with structured analysis tools that don’t require switching into special modes.

The version drawing the most attention is V3.2-Speciale. DeepSeek says it outperformed GPT-5 in the company’s tests and stands toe-to-toe with Gemini 3 Pro on demanding tasks that hinge on logic and multi-step reasoning. As an illustration, the team points to strong results on the 2025 International Mathematical Olympiad and the Informatics Olympiad, adding that the final solutions have been posted publicly.

According to the developers, this leap in capability rests on two pillars: an improved sparse-attention mechanism for handling long contexts, and an expanded reinforcement learning cycle built on more than 85,000 complex problems generated by their multi-step scenario synthesis system.

DeepSeek V3.2 is already available via the company’s website, mobile apps, and API. The more advanced V3.2-Speciale is being released in a limited fashion — it’s running as an experimental reasoning model through a temporary API endpoint that will be shut down after December 15, 2025.

Independent benchmarks still lie ahead, but one trend is already visible: DeepSeek is trying to reset what the industry believes a cutting-edge AI system should cost. If these internal numbers are borne out, that stance is likely to pressure incumbents to rethink their playbooks and focus on what truly drives leadership in this market.