According to a report by the Sci-Tech Daily, DeepSeek is set to release its latest large model - V4 next week. This new model is highly anticipated, featuring capabilities in image, video, and text generation, marking another leap forward for DeepSeek in multimodal technology.

According to insiders, DeepSeek V4 will fully support domestic computing power in hardware compatibility, aiming to optimize compatibility with China-made chips. This not only has the potential to boost demand for its semiconductor products in the Chinese market but also accelerate the integration of AI models with local chips during the inference phase.

At the time of the V4 release, DeepSeek will also launch a brief technical note, with a more detailed technical report expected in one month. These actions demonstrate DeepSeek's commitment to technical transparency and market responsiveness.

Currently, the "V4 Lite" test version is also being tested. This version, known as "sealion-lite," has a context window of 1 million tokens, which is nearly eight times larger than the 128K of the V3 series, theoretically allowing it to process the entire text of "The Three-Body Problem" at once. Notably, V4 Lite features a native multimodal architecture, meaning it integrates text and visual understanding during the pre-training phase, rather than through post-hoc integration.

In terms of parameter scale, V4 Lite is expected to have 200 billion parameters, while the full version of V4 may exceed 1 trillion. Additionally, V4 Lite shows strong performance in generating SVG images, capable of producing high-quality images using just 54 lines of code, indicating improvements in spatial reasoning and structured output.

Looking back at DeepSeek's development journey, the company has continuously innovated in inference capabilities and model efficiency since 2023. Since the release of V2 in 2024, DeepSeek has consistently pursued a balance between comprehensive performance and inference efficiency. By 2025, the V3 series launched by DeepSeek not only established a strong foundation in comprehensive performance but also improved user experience through continuous minor updates.

In summary, while waiting for the release of V4, DeepSeek has always made every effort in technological innovation and the layout of multimodal capabilities. This undoubtedly brings new vitality and possibilities to the field of AI.