DeepSeek Launches V4 Model with Huawei Chip Integration, but Independent Benchmarks Show Mixed Performance Compared to Competitors

Apr 25, 2026, 08:51

Introduction

On April 24, 2025, Chinese AI startup DeepSeek released a preview of its V4 model, the first to be adapted for Huawei's Ascend chips. This marked a shift from its previous reliance on Nvidia hardware. However, independent analysis later showed that the model's performance, while improved, did not beat all domestic and international rivals.

Main Body

DeepSeek, the Beijing-based company that gained global attention in early 2025 for its low-cost AI models, unveiled the V4 Pro preview on April 24. The company claimed that the Pro version outperformed other open-source models on world-knowledge benchmarks, behind only Google's closed-source Gemini-Pro-3.1. A key aspect of the launch was the close collaboration with Huawei. The Chinese telecommunications firm confirmed that its Ascend chips were used in parts of the V4's training process. This shift from Nvidia's AI chips highlights a broader trend in China's semiconductor sector, where domestic alternatives are being prioritized due to US export controls. Nvidia CEO Jensen Huang had previously warned that such a shift could harm the US company's developer ecosystem in China. The V4 model also includes a lower-cost Flash version, and DeepSeek indicated that the preview phase would allow for real-world feedback before a final release, for which no timeline was provided. Shortly after the launch, the benchmarking firm Artificial Analysis released an evaluation on April 25 that placed the V4 Pro second among open-source models, behind Moonshot AI's Kimi K2.6. On the Artificial Analysis Intelligence Index, V4 Pro scored 52, compared to Kimi K2.6's 54. Leading closed-source US models—OpenAI's GPT-5.5, Anthropic's Claude Opus, and Google's Gemini 3.1 Pro—scored 60, 57, and 57, respectively. While the V4 Pro showed a clear improvement over its predecessor V3.2, it did not match the top-tier closed-source systems. AI engineer Daniel Dewhurst, who tested the V4 after its release, cautioned that the benchmark headlines should be treated with caution until independent evaluations and broader developer testing are conducted. He noted, however, that the V4 demonstrated that open models are narrowing the gap with closed models, especially in cost efficiency, long-context handling, and coding tasks. The model can process over one million tokens, comparable to the context windows of OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6, but at a much lower computational cost. Limitations include a lack of support for multiple modalities such as images and video. DeepSeek emphasized that the V4 is particularly suited for AI agent tasks, which require more computing power than standard chatbots. The model quickly rose to the top spot on Hugging Face, according to machine learning engineer Lewis Tunstall. He noted that the V4 excels at handling extremely long and complex text tasks at a lower cost than competing top models. Despite these strengths, the model's inability to process multiple modalities and its reliance on a single benchmark claim have led analysts to call for further independent verification. The launch occurred one day after the White House accused China of large-scale theft of US AI intellectual property, and ahead of a planned visit by US President Donald Trump to Beijing in May. DeepSeek has faced criticism from Washington and US rivals for allegedly relying improperly on American know-how. The company has acknowledged using Nvidia chips but has not clarified whether those chips were subject to export bans, and it has denied intentionally using synthetic data from OpenAI. The Trump administration approved the sale of Nvidia's H200 chips to China in January, but shipments have reportedly been delayed due to disagreements over sales terms in both countries. In financial markets, Chinese chipmakers Huahong Semiconductor and SMIC saw their shares rise 15% and 10%, respectively, on expectations of increased domestic chip adoption. Nvidia's shares also rose, partly due to Intel's positive revenue forecast, which strengthened confidence in sustained AI growth. DeepSeek's own valuation is reported to exceed US$20 billion, with tech giants Alibaba and Tencent reportedly in discussions to take stakes. The V4's release caused shares of domestic rivals Zhipu AI and MiniMax to decline by 9% each.

Conclusion

DeepSeek's V4 model represents a significant technical step forward and a strategic shift toward domestic hardware, but its performance compared to both domestic and US rivals remains debated. The launch highlights China's progress in AI self-sufficiency amid geopolitical tensions, though the model's ultimate success will depend on independent validation and real-world adoption.

Vocabulary Learning

benchmark (n.)
a standard or point of reference against which things may be compared基準；標桿
Example:The model's performance was measured against industry benchmarks.
collaboration (n.)
the action of working with someone to produce something合作；協作
Example:The close collaboration with Huawei was a key aspect of the launch.
evaluation (n.)
the process of judging or assessing the value or quality of something評估；評價
Example:An independent evaluation was released shortly after the launch.
limitation (n.)
a restricting condition or weakness限制；局限
Example:The model's limitations include a lack of support for multiple modalities.
validation (n.)
the act of confirming or verifying something驗證；確認
Example:The model's ultimate success will depend on independent validation.

Sentence Learning

DeepSeek, the Beijing-based company that gained global attention in early 2025 for its low-cost AI models, unveiled the V4 Pro preview on April 24.
Relative clause: The sentence uses a non-defining relative clause 'that gained global attention...' to provide additional information about the company. This helps organize the idea by identifying the subject before describing the action.關係從句：句子使用非限定性關係從句「that gained global attention...」來提供關於公司的額外信息。這有助於在描述行動之前先識別主體，從而組織思路。
The Chinese telecommunications firm confirmed that its Ascend chips were used in parts of the V4's training process.
Passive voice: The phrase 'were used' is passive voice, focusing on the chips rather than who used them. This is common in formal reporting to emphasize the action or result.被動語態：「were used」是被動語態，焦點在芯片而非使用它們的人。這在正式報導中常見，以強調行動或結果。
While the V4 Pro showed a clear improvement over its predecessor V3.2, it did not match the top-tier closed-source systems.
Linking word 'While' for contrast: 'While' introduces a contrast between the improvement and the failure to match top-tier systems. It helps structure a balanced argument.對比連接詞「While」：「While」引出了改進與未能達到頂級系統之間的對比。它有助於構建平衡的論述。
The model can process over one million tokens, comparable to the context windows of OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6, but at a much lower computational cost.
Linking word 'but' for contrast: 'but' contrasts the large token capacity with the lower cost, showing a trade-off. It clearly links two related points.對比連接詞「but」：「but」對比了大的令牌容量與較低成本，顯示了取捨。它清楚地連接了兩個相關要點。
Despite these strengths, the model's inability to process multiple modalities and its reliance on a single benchmark claim have led analysts to call for further independent verification.
Linking word 'Despite' for contrast: 'Despite' introduces a concession (strengths) followed by a contrasting limitation. It organizes the idea by acknowledging positives before negatives.對比連接詞「Despite」：「Despite」引出了一個讓步（優勢），然後是對比的限制。它通過先肯定優點再指出缺點來組織思路。