DeepSeek released 2 models with a breakthrough in agentic systems and AI
Chinese startup DeepSeek released 2 models that claim to be a breakthrough in agentic systems. And judging by the metrics, this is not just marketing.
DeepSeek-V3.2 — this is the official successor to the experimental version. Available in the app, on the website and through API. DeepSeek-V3.2-Speciale — an improved version with emphasis on advanced multi-step reasoning. So far works only through API.
Both models emphasize deep reasoning chains and behavior for agentic scenarios. This is planning, problem solving, complex inferences and work with structured data.
DeepSeek-V3.2-Speciale became the first open-source model that wins gold at top olympiads. Gold at 4 prestigious olympiads! By metrics, Speciale surpasses Gemini 3.0 Pro in mathematics, and the less powerful DeepSeek-V3.2 beats Claude-4.5 Sonnet in coding.
But there’s a nuance. Test-time compute is huge. Speciale doesn’t save tokens at all, so inference turns out expensive. The authors themselves admit they “left optimization for future research”.
Technical reasons for success: this is the new DeepSeek Sparse Attention architecture, large-scale stable RL training and a large pipeline for agentic tasks. And this is the key architecture change compared to the previous generation.
Both models are extremely good at all sorts of agentic tasks, and especially at search and browser tasks. For this, 1800 synthetic environments were generated in which agents trained to perform completely different tasks.
A very cool model turned out, respect.