Chinese startup DeepSeek has introduced a preview edition of its V4 artificial intelligence model, optimized to function on Huawei’s latest Ascend AI chips. The open-source model is presented in two variations: Pro, a top-tier system aimed at delivering competitive performance compared to proprietary competitors, and Flash, a more affordable and quicker alternative. Both versions are equipped to handle a context window of 1 million tokens.
DeepSeek highlighted that V4 is tailored to integrate with agent frameworks like Claude Code and OpenClaw, aligning with the industry trend moving away from prompt-based chatbots towards models capable of performing multi-step tasks with reduced human intervention. In maximum reasoning mode, Pro surpasses all open-source models, as per a paper accompanying the preview. However, it still lags behind leading proprietary systems such as Google’s Gemini 3.1 Pro and OpenAI’s GPT-5.4 in certain aspects.
A notable update from prior DeepSeek releases is the synchronization with Huawei hardware. Following a report from Reuters, Huawei confirmed swift support for V4 on its Ascend 950-based supernode clusters, with its chips utilized in a segment of Flash’s training process.
“Huawei has expanded support for the DeepSeek-V4 series models across its entire Ascend supernode product range through close technical collaboration,” as stated in the Reuters report. DeepSeek did not specify whether V4 underwent training on Nvidia chips, unlike its previous V3 and R1 models that were developed using Nvidia hardware.
DeepSeek mentioned that Pro may incur costs up to 12 times higher than Flash due to limitations in high-end computational capacity, leading to restricted availability at present. According to Reuters, the company anticipates a significant reduction in Pro pricing once Huawei Ascend 950 supernodes are widely deployed in the latter half of the year. DeepSeek is confronted with computational constraints under US export regulations on Nvidia chips and chip manufacturing equipment.
