In the world of AI there is one company that is called whale — DeepSeek. The Chinese company uses a playful whale as its logo and the way it challenged American companies like OpenAI, Anthropic and Google in January 2025, it has earned many fans. Now after weeks of speculation, DeepSeek has finally launched an update to its popular AI model. The update, called DeepSeek V4, takes an aim at the latest Claude, Gemini and ChatGPT. While in some areas it wins some brownie points, in others it is a miss for the blue whale.
One area where the DeepSeek 4 succeeds is in reducing the cost of AI, and hence offering a one-million context window. In other words, you can put one million tokens (made up of words or characters) in the AI model and it will be able to generate output on the basis of it all. Doing so with American AI models is either not possible or terribly expensive.
With support for a massive context window, DeepSeek says its new models can handle long tasks, complex reasoning, and AI agents in an easier and more practical manner.
The new V4 lineup from DeepSeek includes two models: DeepSeek-V4-Pro and DeepSeek-V4-Flash. The Pro version comes with 1.6 trillion total parameters, though only 49 billion are active at a time of use.The smaller Flash model carries 284 billion total parameters with 13 billion active. Both models have been trained on more than 32 trillion tokens and are said to be built to process huge amounts of text in a single go.
DeepSeek claims the V4 series can compete with leading proprietary models like Claude, GPT and Gemini across several benchmarks. The company highlights that, in its own results, the highest reasoning configuration — DeepSeek-V4-Pro-Max — performs strongly in coding and reasoning tasks, even surpassing rivals in areas such as coding benchmarks and shortlist evaluations. But at the same time, the wins are not clean for DeepSeek. In many tasks, the model trails top closed AI tools like Claude and ChatGPT in certain knowledge-heavy tests.
Instead of outright performance, DeepSeek is trying to stand out in efficiency compared to the US AI models. With its V4 model, the company has introduced a hybrid attention system that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA).
In simple terms, these techniques reduce how much data the model needs to process at any given moment, especially when handling very long inputs. The result is a significant drop in compute requirements. According to the paper released by the company, DeepSeek-V4-Pro uses just 27 per cent of the compute and 10 per cent of the memory cache compared to its previous V3.2 model at the same one-million-token setting, while the Flash variant pushes efficiency even further.
China vs US AI race continues
While DeepSeek V4 seems like a major AI release in terms of its underlying technology — which is much more efficient than current AI models — the outright performance shows that Chinese AI is still around half a year behind the US models. In most benchmarks, DeepSeek V4 matches ChatGPT 5.2 or Claude 4.6, which are now one or two generations behind the latest AI models.
However, yet again DeepSeek is trying to upset the American models through its clever open-source play. Just the way it matched the top American AI models last year in January, it is again matching some of the best AI models but with an open-source model.
The new Chinese AI model is also available at a much cheaper price through API to companies across the world, and DeepSeek must be hoping that instead of paying top dollars to something like Claude 4.7, which is excellent but pricey, organisation would instead pay for DeepSeek V4 that is good enough and relatively dirt-cheap as far as AI tools are concerned.
