DeepSeek rolled out the new V4.1-Flash. The company is once again betting on low pricing: the model costs less than the previous version while delivering results on par with more expensive AI models.
What V4.1-Flash can do
-
Less VRAM. Running the model now requires roughly four times less memory than before. This is good news for those who run AI on their own GPUs.
-
Less disk space. The model size has been reduced by about eight times, so it takes up much less space.
-
Understands images. An image can be attached to a request via the API. This will be useful for services, bots, and other projects where AI needs to analyze more than just text.
-
Cheaper at night. Prices are twice as low during off-peak hours. One million input tokens with a cache hit costs $0.003, without cache — $0.15, while one million output tokens costs $0.60. During the day, prices are twice as high.
What about quality?
DeepSeek claims that V4.1-Flash outperforms the more expensive V4 Pro in a number of tests. The model performs especially well in programming, logic tasks, and AI agent workloads.
At the same time, it is too early to immediately call it the best model on the market: some of the comparisons were conducted by DeepSeek itself.
It looks like DeepSeek has once again decided to start a price war in the AI market.
