DeepSeek V4: Huawei Replaces NVIDIA, Challenging GPT-5.5

The world of artificial intelligence has been shaken by the announcement of DeepSeek V4, a model from the Chinese startup DeepSeek, based in Hangzhou. This model was launched on April 24, just one day after the presentation of GPT-5.5 by OpenAI on April 23. What makes DeepSeek V4 particularly noteworthy is its exclusive use of Huawei's Ascend chips, without any reliance on NVIDIA GPUs.

Two Variants, Two Approaches

DeepSeek has chosen to release its V4 model in two distinct versions: Pro and Flash. The Pro version is a true behemoth with 1.6 trillion parameters, but thanks to the Mixture-of-Experts architecture, only 49 billion parameters are activated for each query. Meanwhile, the Flash version is more compact, with 284 billion parameters, of which 13 billion are activated. Both versions share a context window of one million tokens.

The Pro version is designed to handle complex tasks and advanced agentic capabilities, while the Flash version is optimized for fast and economical inference. Both models are accessible via the DeepSeek website, in "Instant" or "Expert" mode, as well as through API.

A Complex and Delayed Development

The launch schedule for DeepSeek V4 has been particularly hectic. Initially planned for February, then March, the launch was delayed due to the need to migrate the entire training code from NVIDIA's CUDA ecosystem to Huawei's in-house framework, CANN. This complex process ultimately succeeded, allowing for the release of this ambitious model. An intermediate version, V4-Lite, had even briefly surfaced on March 9.

A Pioneering Model on Chinese Silicon

DeepSeek V4 is the first frontier-class AI model fully trained on Chinese chips, marking a significant break from the traditional reliance on NVIDIA GPUs. Until now, even the most advanced labs in Beijing depended on NVIDIA GPUs for their heavy models. Furthermore, DeepSeek chose not to pass its V4 model to NVIDIA or AMD for optimization, a common practice in the industry, preferring to offer this opportunity to Chinese manufacturers.

The timing of the DeepSeek V4 launch is also significant. It came just after the release of a memo from the White House on April 23, accusing China of "industrial theft by distillation." By releasing an open-source model so quickly after GPT-5.5, DeepSeek has reduced OpenAI's media monopoly to just a few hours, highlighting a new era of technological competition where hardware independence becomes a major strategic asset.