DeepSeek expands the V4 family with V4 Flash Vision Exp: a ‘smaller’ AI model that is on par with Anthropic Claude Opus 4.8 in terms of capabilities.
The Chinese company DeepSeek is adding a new model to the V4 series launched in April. V4 Flash Vision Exp is based on the V4 Flash model and reportedly scores better in almost all benchmarks. For now, the model is only available via the API platform for Chinese users, but DeepSeek usually makes its models available as open source.
Via Hugging Face, we find more technical information about the model. V4 Flash Vision Exp has a total of 284 billion parameters, evenly distributed across twenty ‘smaller’ neural networks of thirteen billion parameters. The total number of parameters is five times lower than the flagship V4 Pro model. V4 Flash Vision Exp is said to perform much better than V4 Flash, primarily in image generation.
Benchmarks compare the new DeepSeek model with what the competition has to offer. In terms of capabilities, V4 Flash Vision Exp is closest to Anthropic’s Claude Opus 4.8. That model sits below Claude Opus 5 and Fable 5 in Anthropic’s hierarchy.
Bigger is not better
Recent developments in AI adoption by companies show that you do not always have to launch the largest, most powerful model to be successful. Companies are now looking specifically for smaller, specialized models to curb costs. Cheaper Chinese models are suddenly gaining relevance, and DeepSeek is symbolic of this trend.
Anthropic is finding it more difficult to market its best models, even to business users. The company itself is convinced that businesses are willing to pay top dollar for the highest quality, but reality seems to contradict this.
Yet there is also a dark side to the breakthrough of Chinese AI technology. They are often developed using distillation, a non-prohibited yet contested technique for training models. The use of Chinese technology is also a sensitive issue for privacy and geopolitical reasons. We delve deeper into this in this article.
