DeepSeek has released a preview of V4, its new flagship open-source model and its biggest launch since R1 in January 2025. The new system can process much longer prompts than the previous generation and is available for anyone to download, use, and modify. The release comes after months of scrutiny around personnel changes, delayed launches, and pressure from both the US and Chinese governments, but it still stands out as an important step for open models and for China’s broader Artificial Intelligence ambitions.
V4’s first significance is its position in the open-source market. DeepSeek says the model rivals top closed-source systems while remaining much cheaper to use, giving developers and companies access to frontier capabilities without the same cost burden. The new family includes V4-Pro for coding and complex agent tasks and V4-Flash for faster, lower-cost deployment, with both versions supporting reasoning modes. Benchmark results shared by the company place V4-Pro alongside leading systems from Anthropic, OpenAI, and Google, while ahead of open-source rivals such as Alibaba’s Qwen-3.5 and Z.ai’s GLM-5.1 on coding, math, and STEM tasks. In an internal survey of 85 experienced developers, More than 90% included V4-Pro among their top model choices for coding tasks.
The second major development is a new approach to long-context memory efficiency. Both versions can handle 1 million tokens, which DeepSeek says is now the default across its services. The model achieves this through changes to its attention mechanism, compressing older information and focusing more selectively on what matters while preserving nearby text in full detail. In a 1-million-token context, V4-Pro uses only 27% of the computing power required by its previous model, V3.2, while cutting memory use to 10%. The reduction in V4-Flash is even larger, using just 10% of the computing power and 7% of the memory. That could make large-context applications more practical, including coding assistants that read entire codebases or research agents that analyze long archives without losing track of earlier material.
The third reason V4 matters is hardware strategy. V4 is DeepSeek’s first model optimized for domestic Chinese chips such as Huawei’s Ascend, making the launch an early test of whether China can reduce dependence on Nvidia. Huawei said its Ascend supernode products, based on the Ascend 950 series, would support DeepSeek V4, and DeepSeek says V4-Pro prices could fall significantly after Huawei’s Ascend 950 supernodes begin shipping at scale in the second half of this year. DeepSeek has not fully moved beyond Nvidia, however. Its technical report indicates Chinese chips are being used for inference, while questions remain about how much of the training process, including key long-context features, has been adapted to domestic hardware. Even so, the release suggests China is making progress toward a more self-reliant Artificial Intelligence stack.
