DeepSeek just released V3.2 — a new iteration of the much loved open source large language model. What improved compared to V3.1 and where does the new version sit relative to flagships from OpenAI, Gemini, and Claude, and is it the new best AI model?
The new model is already live on Overchat AI, and you can start chatting with DeepSeek V3.2 today.
Here's everything you need to know about the new model.
DeepSeek V3.2 is a flagship-level open-source AI model released in December 2025 by the Chinese AI research lab DeepSeek.
For the first time, the model introduces DeepSeek Sparse Attention (DSA), a new training approach that greatly improves the efficiency of long-context training and inference, essentially enabling DeepSeek to develop a competitive flagship model at a fraction of the cost.
Training AI models is notoriously hard and expensive, and as a result, while many top AI labs sell their strongest models for high prices, DeepSeek continues to release powerful models as open source. According to the team, V3.2 delivers performance similar to GPT-5 for everyday tasks.
The model comes in two configurations:
New to DeepSeek V3.2 is its ability to think while using internal tools. This means the model can mimic human reasoning and use tools like search engines and calculators as needed.
To put that into context, the DeepSeek V3.2-Speciale model excels at math, achieving scores of 96.0 on the AIME 2025 and 99.2 on the HMMT Feb 2025. These scores surpass those of the GPT-5 High and Gemini 3.0 Pro models. Currently, the DeepSeek V3.2-Speciale is the most powerful model for math. However, each of these models is virtually perfect for consumer use, including college- and university-level applications. .
On LMArena leaderboards, both DeepSeek models currently lag behind GPT-5 and Gemini 3.
For coding, Gemini 3.0 Pro is still ahead, but DeepSeek V3.2-Speciale comes close with competitive scores of 88.7 and 2701 on LiveCodeBench and Codeforces, respectively.
Keep in mind that benchmark performance doesn't capture everything about real-world use. A model with slightly lower benchmark scores might perform better with good prompting than a technically stronger model used poorly.
In summary, it’s impressive that DeepSeek was able to train a model that rivals the leading western closed-source models. Notably, unlike western models, it’s open-source and requires less processing power.
Let's compare DeepSeek V3.2 against the competition.
The DeepSeek V3.2 performs similarly to the GPT-5.1 model for everyday tasks, but it is more powerful when it comes to math.
In terms of pure mathematical reasoning, DeepSeek V3.2-Speciale often outperforms GPT-5, particularly on specialized benchmarks like HMMT. However, for general conversation and multimodal tasks, GPT-5 still offers a more complete experience.
Gemini 3.0 Pro is Google's most advanced reasoning model. In benchmarks, it outperforms DeepSeek in most categories. Surprisingly, DeepSeek wins in two tests.
To explain these numbers, Gemini 3.0 Pro outperforms DeepSeek in general reasoning tests and is a slightly more powerful tool for general AI use, while DeepSeek V3.2-Speciale is stronger at math.
👉 Chat with Gemini 3.0 Pro on Overchat AI
The fastest way to try DeepSeek V3.2 is on Overchat AI. Chat with DeepSeek V3.2 on Overchat AI
DeepSeek Platform
DeepSeek V3.2 is available directly on DeepSeek's website and through their mobile apps. The standard model is free to use.
DeepSeek V3.2-Special is currently only accessible via the DeepSeek API. API access is on a pay-per-use basis. Though, DeepSeek is less expensive than closed-source models.
For Developers
Developers can access DeepSeek V3.2 via the DeepSeek API. DeepSeek V3.2-Speciale costs 42 cents per million tokens, making it 10 times cheaper than flagship models from closed-source competitors.
DeepSeek V3.2 shows that the Chinese AI lab is still holding its own against major players like Google and OpenAI, continuing to push the limits of training large, high-performance models.
It delivers performance on par with GPT-5 and Gemini 3.0 Pro — especially in math and reasoning—yet costs only a fraction as much. That makes it a strong everyday model for general use and a highly dependable choice for math, coding, and more advanced reasoning tasks.
👉 Start chatting with DeepSeek V3.2 on Overchat AI today
DeepSeek V3.2, an advanced open-source AI model, was released in December 2025 by DeepSeek, a Chinese AI research lab. It’s the first model to introduce sparse attention mechanisms, which allowed DeepSeek to train a model that rivals Western closed-source flagship-level AI while costing 10x less to run.
DeepSeek V3.2 was released in December 2025, with the experimental version (V3.2-Exp) announced earlier that month.
DeepSeek V3.2 outperforms GPT-5 in math, reasoning, and other specialized problem-solving tasks, while costing about ten times less to run than ChatGPT. It’s also fully open source, which makes its strong performance even more impressive compared to GPT-5’s closed-source approach.
ChatGPT, however, still offers a more polished user experience and leads in image and especially video capabilities, thanks to Sora 2.
Gemini 3.0 Pro, released in November 2025, quickly surpassed every other available model on most benchmarks, especially in coding and reasoning. It’s the most advanced model Google has ever produced.
With that in mind, it’s especially impressive that DeepSeek V3.2 beats Gemini 3.0 Pro in math on the HMMT benchmark and stays neck-and-neck on most others. In practice, DeepSeek V3.2 delivers Gemini 3.0 Pro–level performance at just a fraction of the cost.
You can try DeepSeek V3.2 on Overchat AI right now. You can also access them through DeepSeek's website, mobile apps, and API.
