DeepSeek V4 Emerges as Multimodal AI Game-Changer

Sci-Tech Daily reports that DeepSeek will launch its newest large-scale model, V4, next week. This highly anticipated release introduces image, video, and text generation capabilities, representing a major step in DeepSeek’s multimodal technology development.
Insiders indicate that DeepSeek V4 will fully support domestic computing hardware, focusing on enhanced compatibility with China-made chips. This approach could increase semiconductor demand in the local market and accelerate integration between AI models and domestic chips during inference.
Alongside the V4 launch, DeepSeek will publish a concise technical note, with a more comprehensive report expected in one month. These steps highlight the company’s dedication to transparency and responsiveness to market needs.
Currently, the “V4 Lite” test version, named “sealion-lite,” is undergoing trials. It supports a 1-million-token context window—nearly eight times larger than the V3 series’ 128K—enabling it to process the entire text of “The Three-Body Problem” in one go. Importantly, V4 Lite is built with a native multimodal architecture, integrating text and visual understanding from pre-training rather than through later modifications.
In terms of parameters, V4 Lite is projected to include 200 billion, while the full V4 version may surpass 1 trillion. V4 Lite also excels in SVG image generation, producing high-quality visuals with as few as 54 lines of code, signaling progress in spatial reasoning and structured output.
Reflecting on DeepSeek’s growth, the firm has consistently advanced inference capabilities and model efficiency since 2023. Starting with V2 in 2024, DeepSeek has emphasized a balance between overall performance and inference efficiency. By 2025, the V3 series solidified this foundation and enhanced user experience through incremental updates.
In summary, as the AI community anticipates V4, DeepSeek continues to drive innovation and expand multimodal capabilities, injecting fresh energy and possibilities into the AI landscape.
Related article
AI Search Mandatory Policy Fuels Exodus, DuckDuckGo Sees User Surge
Following Google's 2026 I/O conference announcement of a full AI overhaul of its search engine, many users started looking for more controllable alternatives because there was no simple "one-click disable" for AI features. The privacy-focused search
Xiaohongshu Restructures: Conan Named President, Creates AI Primary Department Dots and Overseas Division Rednote
On April 30, Xiaohongshu sent an internal memo to all employees announcing the launch of a new organizational restructuring. The core of this change involves fully integrating three business lines—community, e-commerce, and commercialization—along wi
Tencent's Xiaolongxia Surges Beyond Expectations, Team Expands Capacity 10x, Apologizes and Compensates
Tencent has officially launched WorkBuddy, an all-scenario AI intelligent agent, marking a new phase in the large model application layer race with high integration and a low deployment threshold.The product drew immediate industry attention on its l
Related Special Topic Recommendations
Comments (0)
0/500

Sci-Tech Daily reports that DeepSeek will launch its newest large-scale model, V4, next week. This highly anticipated release introduces image, video, and text generation capabilities, representing a major step in DeepSeek’s multimodal technology development.
Insiders indicate that DeepSeek V4 will fully support domestic computing hardware, focusing on enhanced compatibility with China-made chips. This approach could increase semiconductor demand in the local market and accelerate integration between AI models and domestic chips during inference.
Alongside the V4 launch, DeepSeek will publish a concise technical note, with a more comprehensive report expected in one month. These steps highlight the company’s dedication to transparency and responsiveness to market needs.
Currently, the “V4 Lite” test version, named “sealion-lite,” is undergoing trials. It supports a 1-million-token context window—nearly eight times larger than the V3 series’ 128K—enabling it to process the entire text of “The Three-Body Problem” in one go. Importantly, V4 Lite is built with a native multimodal architecture, integrating text and visual understanding from pre-training rather than through later modifications.
In terms of parameters, V4 Lite is projected to include 200 billion, while the full V4 version may surpass 1 trillion. V4 Lite also excels in SVG image generation, producing high-quality visuals with as few as 54 lines of code, signaling progress in spatial reasoning and structured output.
Reflecting on DeepSeek’s growth, the firm has consistently advanced inference capabilities and model efficiency since 2023. Starting with V2 in 2024, DeepSeek has emphasized a balance between overall performance and inference efficiency. By 2025, the V3 series solidified this foundation and enhanced user experience through incremental updates.
In summary, as the AI community anticipates V4, DeepSeek continues to drive innovation and expand multimodal capabilities, injecting fresh energy and possibilities into the AI landscape.
AI Search Mandatory Policy Fuels Exodus, DuckDuckGo Sees User Surge
Following Google's 2026 I/O conference announcement of a full AI overhaul of its search engine, many users started looking for more controllable alternatives because there was no simple "one-click disable" for AI features. The privacy-focused search
Xiaohongshu Restructures: Conan Named President, Creates AI Primary Department Dots and Overseas Division Rednote
On April 30, Xiaohongshu sent an internal memo to all employees announcing the launch of a new organizational restructuring. The core of this change involves fully integrating three business lines—community, e-commerce, and commercialization—along wi
Tencent's Xiaolongxia Surges Beyond Expectations, Team Expands Capacity 10x, Apologizes and Compensates
Tencent has officially launched WorkBuddy, an all-scenario AI intelligent agent, marking a new phase in the large model application layer race with high integration and a low deployment threshold.The product drew immediate industry attention on its l





Home






