Despite a measurable step up from its predecessor, DeepSeek’s V4 Pro still trails the front-runners in the latest benchmarks. The model improved on V3.2, but posted a score of 52 on the Artificial Analysis Intelligence Index – behind Moonshot AI’s Kimi K2.6, which reached 54 shortly after its release this week. 

The gap becomes more pronounced against leading US closed-source systems, with OpenAI’s GPT-5.5 scoring 60, while Anthropic’s Claude Opus and Google’s Gemini 3.1 Pro both registered 57. 

The latest results also indicate a mounting pressure on DeepSeek as China pushes to close the AI capability gap with the US, while dealing with a fierce domestic rivalry and persistent compute limitations.

Balancing performance with hardware flexibility

The V4 system is considered impressive for approaching state-of-the-art performance while maintaining efficiency, combining a one-million-token context window with compatibility for Huawei Technologies’ Ascend 950PR AI chips. According to Kyle Chan of the Brookings Institution, this balance of performance and hardware adaptability signals meaningful progress, even if the model does not yet surpass top competitors, the South China Morning Post writes.

In practical terms, an AI model’s context window defines how much information it can handle in a single pass. DeepSeek’s earlier flagship system supported up to 128,000 tokens, but the new architecture significantly expands that capacity while improving efficiency. 

According to a recent report by SemiAnalysis, DeepSeek has achieved a roughly 90% reduction in KV cache usage when operating within a one-million-token context. The firm described this optimization as more impactful than the TurboQuant paper published last month by Google, underscoring the importance of memory efficiency as models scale to longer contexts.

Another key aspect of the release is its alignment with China’s domestic hardware ecosystem. Soon after V4 debuted, Huawei Technologies confirmed that its Ascend chip lineup and supernode systems would fully support the model at the inference stage. However, the training side remains less transparent. Experts also noted that DeepSeek made no reference to using Chinese chips during training, which is a notable omission given the broader push for technological self-sufficiency, and one that stands out as the model continues to lag behind leading US frontier systems.

New model leaves global AI balance largely unchanged 

From a geopolitical standpoint, the latest release has not materially altered perceptions of US dominance in AI. Chris McGuire of the Council on Foreign Relations said the gap remains intact, estimating the US is still roughly seven months ahead. He also pointed to the lack of disclosure around training costs and infrastructure, suggesting that V4 may have depended on restricted Nvidia Blackwell chips, raising further questions about how the model was developed under tightening export controls.

Investor response to the launch was notably restrained compared with the shockwaves triggered by DeepSeek’s earlier R1 model. While news of V4 and its integration with Huawei Technologies hardware lifted shares of Chinese chipmakers on Friday, the broader global market reaction remained subdued. 

By contrast, the R1 debut erased hundreds of billions of dollars from US equities, with Nvidia plunging 17% in a single session. This time, sentiment moved in the opposite direction as Nvidia stock rose 4.32%, showing how V4 has yet to deliver a comparable market jolt. Even with stronger performance on knowledge benchmarks, reliability remains a concern for DeepSeek’s latest models.