A Chinese company called DeepSeek has put out a new artificial intelligence model. They call it DeepSeek-V4, and they are making some pretty bold claims about how good it is.
Basically, the company says this new model beats every other open source competitor on the market.
DeepSeek, which is based in Hangzhou, says the model is especially good at handling very large amounts of text. It can process up to one million words in one go. To put that in perspective, it can read and understand entire books or massive databases.
The model comes in two flavors: a Pro version and a Flash version. The Flash version is the cheaper and more efficient option. And as the name suggests, the Pro version is the powerful one. According to DeepSeek, the Pro variant does better than other open source models on world knowledge tests. The DeepSeek-V4-Pro also has a special mode that uses extra reasoning power.
The only model that still beats it by a small margin is Google’s Gemini-Pro-3.1, which is not open source.
With the ongoing AI race, the United States has placed more restrictions on selling advanced computer chips to China. That’s why Chinese companies are having to rely more on homegrown chips.
DeepSeek has not said exactly which chips they used to train the new V4 model. But they did say their software works with both Nvidia and Huawei chips.
One of the biggest improvements in V4 is speed and memory. The model can process up to 384,000 tokens, compared to the older V3 model 128,000 tokens limit.