DeepSeek’s V3.2 & V3.2-Speciale: A new challenger to GPT-5 and Gemini 3 Pro

Chinese AI company with claDeepSeek proposed two new large language models, DeepSeek-V3.2 and DeepSeek-V3.2-Specialeimed performance roughly equal to or better than the recent proprietary frontier models (in both OpenAI and Google DeepMind), and which are open-source and more compute-efficient. The launch has excited the AI community in the world, particularly among people willing to develop powerful, yet affordable models.

What Are V3.2 and V3.2-Speciale?

DeepSeek characterises the novel models as developed based on the "Mixture-of-Experts" transformer architecture, which has approximately 671 billion total parameters. Nevertheless, active parameters are only a few, of the order 37 billion, per token when performing inference - a typical MoE approach to cutting down costs without changing capacity.

The most important technical advancement is a new model known as DeepSeek Sparse Attention (DSA). DSA aims to simplify the complexity of computation - especially the long input contexts - by dividing attention into two parts: a lightweight "selector/indexer" that filters tokens of interest, and then densely attending to them. This enables the model to process very long contexts more effectively than the conventional dense-attention LLMs.

In addition to sparse attention, DeepSeek refers to a Scalable Reinforcement Learning Framework and a Large-Scale Agentic Task Synthesis Pipeline - suggesting that the models were not only trained on passive text data but also on synthetic tasks (e.g. tool-use, code generation, complex reasoning), to increase their action the ability of reason and execution of multi-step workflows.

DeepSeek claims that such improvements make V3.2 a balanced inference - model - because it can be used every day, and V3.2-Speciale is the maxed-out reasoning kind, which is used in high-demand tasks (such as math reasoning, code writing, and agentic loads).

How DeepSeek’s V3.2 & V3.2-Speciale perform: Benchmark claims & public results

DeepSeek has released benchmark scores which assert V3.2-Speciale scores best-in-class on difficult math, reasoning and coding exams - even in cases where they perform better than the best proprietary models.

• A recent article in VentureBeat reports, V3.2-Speciale is above GPT-5-High and Gemini 3 Pro with a pass rate of 96.0% on the AIME 2025 math benchmark. The model was reported to have a score of around 99.2 on the Harvard-MIT Mathematics Tournament (HMMT).

• DeepSeek and media news reports that V3.2-Speciale performed with gold-medal performance on the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI), and well on competitive programming contests (e.g. ICPC).

DeepSeek also performs at a competitive level in the area of coding (e.g. bug fixing or process sequences). The company provides that the open-source models perform strongly with less computational resources needed as compared to most proprietary models.

Assuming that they are independently tested, these findings, particularly on math reasoning and coding, would be a milestone: a genuinely open-source model that could compete (or even surpass) closed-source frontier systems, at one-tenth the price.

DeepSeek’s V3.2 & V3.2-Speciale: Caution & independent verification needed

Nonetheless, regardless of the outrageous assertions and methodological advances, there are significant cautions.

• Results of the performance are calculated on the basis of benchmark reports submitted (or at least published) by DeepSeek or friendly media sources. Publicly accessible independent peer-reviewed assessments are still not available.

• Certain outcomes (primarily those that make references to math-Olympiad performance or ICPC performance) are too good to be true. In the past, the performance of LLMs on such competitive exams also tended to perform poorly under realistic conditions (time constraints, real-world problem description, human-like reasoning, robustness, etc.). The external validation will play a significant role.

• Not widely open-sourced (only through API) yet, the "Speciale" model, the highest-performance one, may not allow extensive experimentation yet until weights become accessible or otherwise.

• Sparse attention is real, and running large models (millions of tokens, long-context reasoning, agentic tasks) still might need a lot of compute - which is potentially out of reach of resource-constrained users.

So the promise is gigantic, but these claims are only made as the game-changing things under consideration, not facts.

A turning point in Open-Source AI

DeepSeek-V3.2 and V3.2-Speciale is one of the most ambitious attempts to date to democratise frontier-level AI: a combination of high-performance reasoning, code and tool-use capabilities, open-source availability, and computational efficiency.

Once the benchmarks can qualify an independent examination, and the community starts to construct upon these paradigms, we might be observing a significant change: the type of elite, closed-source frontier models in favor of open, accessible AI infrastructure a change that has far-reaching consequences on the innovation process and equity, as well as AI uptake worldwide.

This may be the time to listen to developers, researchers, and AI-curious readers, in particular, in those parts of the world where it is expensive.