DeepSeek V4 is here, setting up a showdown with Nvidia

Chinese AI company DeepSeek released a preview version of its new model V4 on April 24, 2026. It is open source, so others can freely download it and build on it. The model comes in two versions: V4-Pro and V4-Flash.

Large models with long memory

According to DeepSeek, V4-Pro has 1.6 trillion parameters in total, of which 49 billion are active at a time. Parameters are the settings an AI model has learned, and the more there are, the larger the model. V4-Flash is smaller, with 284 billion parameters in total and 13 billion active.

Both can work with up to 1 million tokens at a time. A token is a small piece of text, so that corresponds to very long documents or conversations. The models have both a thinking and a non-thinking mode and can be used with the same programming interfaces as OpenAI and Anthropic.

DeepSeek says a new attention mechanism called DSA delivers significantly lower costs for compute and memory on long texts. According to Business Today, V4-Pro uses 27 percent of its predecessor's compute on long texts, and V4-Flash uses 10 percent.

The older models deepseek-chat and deepseek-reasoner will be shut down completely after July 24, 2026.

Price and performance

For companies that use the model via API, V4-Pro costs $1.74 per million input tokens and $3.48 per million output tokens. V4-Flash costs $0.14 per million input tokens.

According to Bloomberg's reporting, V4 outperforms other open source models on knowledge tests and trails only Google's Gemini-Pro-3.1. Business Today writes that V4-Pro competes with GPT-5.4 and Gemini 3.1 and beats open source models in coding, math and the sciences. These are claims that have not been independently tested.

Huawei at the center

The big story is the chips. According to Business Today, V4 is adapted to Huawei's Ascend chips instead of Nvidia's GPUs, which earlier models such as V3 and R1 were trained on. Huawei has announced that it fully supports V4 across Ascend chips. Bloomberg writes that Huawei chips were part of the training. DeepSeek's own announcement does not mention Huawei.

On the stock market, Huahong Semiconductor rose 15 percent and SMIC 10 percent after the launch, while the Chinese AI companies Zhipu AI and MiniMax fell 9 percent each.

According to Bloomberg, Nvidia CEO Jensen Huang has called it a dreadful outcome the day DeepSeek first comes out on Huawei. According to Business Today, he has also said that China will become superior to the US if future AI models are optimized differently from the American technology stack.

Bloomberg also writes that DeepSeek is aiming for a valuation of $20 billion.