[DigitalToday reporter Jinju Hong (홍진주)] Chinese artificial intelligence company DeepSeek unveiled its next flagship model, V4, stepping up its push into the open-source AI market.

MIT Technology Review reported on April 24 that V4 sharply expands long-form processing and boosts price competitiveness, while also highlighting optimisation for Chinese-made semiconductors.

The update marks the biggest change since the R1 reasoning model released in January 2025. DeepSeek unveiled two versions: V4-Pro and V4-Flash. V4-Pro targets coding and complex agent tasks, while V4-Flash is a lightweight model focused on speed and cost efficiency. Both models are available on the web and in apps, and the company also provides an API for developers.

The company touted both price and performance as key strengths. DeepSeek said V4-Pro costs $1.74 per 1 million input tokens and $3.48 per 1 million output tokens. V4-Flash is cheaper at about $0.14 per 1 million input tokens and about $0.28 per 1 million output tokens. The company stressed it is among the cheapest in the top-tier model group.

DeepSeek also expressed confidence in performance competition. It claimed that, based on its own benchmarks, V4-Pro is competitive with Anthropic's Claude series, OpenAI's GPT line and Google's Gemini-3.1 model. It added that, among open-source models, it outperformed Alibaba's Qwen and Z.ai's GLM family in coding, math and STEM fields.

Another change in the model is long-context handling. V4 supports up to 1 million tokens of context and applies structural improvements rather than a simple expansion. It boosts efficiency by compressing older information and selectively focusing on currently needed information. The company said this can maintain accuracy while lowering costs in long-document analysis or large-scale code processing.

The approach is also linked to cost savings. DeepSeek said V4-Pro cuts computing resources by about 27% versus its previous V3.2 model, while reducing memory to about 10% of the prior level. V4-Flash runs on even fewer resources. Such efficiency is expected to help lower the cost of building large-scale AI services.

On the hardware side, the release is drawing attention as a first step toward reducing reliance on Nvidia. V4 is the first model optimised for Huawei's Ascend chips. Huawei said systems based on its Ascend 950 support V4, a move that is interpreted as aligning with China's push for self-reliance in AI infrastructure.

Some analysis says it is not a full transition. The industry views Chinese chips as currently used mainly for inference, and sees a strong possibility that reliance on Nvidia chips still remains in training.

There is also room for further price cuts. DeepSeek said V4-Pro pricing could fall further if Huawei Ascend chips are supplied in large volumes. This is interpreted as a signal of China's AI strategy to build chips, models and infrastructure using domestic technology over the long term.

DeepSeek, which has drawn attention amid recent talent departures and regulatory pressure, put both its technical competitiveness and the potential to expand its ecosystem to the test with V4. The model's market impact is expected to be gauged not only through performance competition, but also within the flow of open-source AI and the global contest for leadership in semiconductors.