Quick Read
- DeepSeek is reportedly launching its V4 large AI model this week, according to the Financial Times.
- DeepSeek V4 is a multimodal model capable of generating images, videos, and text.
- The model features a native multimodal architecture, integrating text and visual understanding from pre-training.
- V4 Lite, a simplified version, is rumored to have a 1 million token context window and 200 billion parameters.
- DeepSeek V4 is optimized for and supports domestic computing power, boosting China’s semiconductor demand.
Reports from informed sources indicate that DeepSeek, a prominent player in artificial intelligence development, is poised to launch its latest large AI model, V4, as early as this week. The Financial Times and STAR Market Daily have revealed that this new iteration is a multimodal model, capable of generating images, videos, and text, marking a significant advancement in the company’s offerings and potentially influencing the broader AI landscape.
The anticipated launch comes at a time of heightened global interest in advanced AI capabilities, particularly those that integrate diverse data types. DeepSeek V4’s expected features, including its native multimodal architecture and extensive support for domestic computing power, position it as a key development for both technological innovation and strategic localization within the AI sector.
DeepSeek V4’s Multimodal Capabilities and Architecture
DeepSeek V4 is expected to feature a native multimodal architecture, meaning its ability to understand and generate content across text and visual mediums is integrated from the foundational pre-training stage. This approach contrasts with models that add multimodal capabilities through post-hoc concatenation, suggesting a more seamless and sophisticated understanding of diverse data inputs. Informed sources suggest that a simplified version, known as V4 Lite and codenamed ‘sealion-lite,’ is currently undergoing secret testing. This Lite version is rumored to boast a context window of 1 million tokens, a substantial increase nearly eight times larger than the 128K tokens supported by the previous V3 series. This expanded capacity would theoretically allow it to process extensive documents, such as entire novels, in a single operation.
While the V4 Lite’s parameter scale is estimated at 200 billion, speculation suggests that the full version of DeepSeek V4 could potentially exceed one trillion parameters, placing it among the most powerful AI models developed to date. Early leaked test examples of V4 Lite demonstrate its capability to generate high-quality SVG images using minimal code, outperforming models like DeepSeek V3.2 and Claude Opus 4.6 in code optimization and visual fidelity. This indicates enhanced spatial reasoning and structured output capabilities, vital for complex creative and technical applications.
Boosting Domestic AI Computing Power
A crucial aspect of DeepSeek V4’s development is its commitment to supporting domestic computing power. According to reports, DeepSeek provides extensive support for chips manufactured in China and is actively optimizing the V4 model to align with these domestic hardware capabilities. This strategic move is expected to significantly boost demand for Chinese semiconductor products within the domestic market and accelerate the localization of AI model ‘inference’ processes.
This focus on domestic hardware compatibility aligns with broader national strategies to foster self-sufficiency in critical technological sectors. By developing models optimized for local infrastructure, DeepSeek aims to reduce reliance on foreign technology, enhance data security, and potentially lower the operational costs for AI deployment within China.
Evolution of DeepSeek’s AI Model Lineup
DeepSeek’s forthcoming V4 model builds upon a robust history of iterative development. The company has maintained two distinct technological approaches: the V series, designed as an ‘all-purpose assistant’ focused on comprehensive performance, and the R series, which serves as a ‘problem-solving expert’ specializing in complex reasoning tasks. The last major update prior to V4 was the release of R1 in January 2025, which was the company’s first dedicated reasoning model.
Key milestones in the V series include the May 2024 release of DeepSeek V2, which introduced the Multi-Head Latent Attention (MLA) mechanism to reduce memory usage and inference costs. This was followed by V3 in December 2024, establishing an efficient Mixed Expert (MoE) architecture for comprehensive performance. Subsequent refinements, such as DeepSeek-V3.1 (August 2025), integrated the capabilities of both V3 and R1, enabling seamless switching between thinking and non-thinking modes and achieving over 40% improvement on benchmarks like SWE-bench. The V3.2 version, launched in December 2025, further advanced the lineup by integrating deep reasoning with tool use, reportedly achieving reasoning capabilities comparable to GPT-5 levels while significantly reducing inference costs on H800 clusters.
The imminent launch of DeepSeek V4, with its native multimodal architecture and strategic alignment with domestic computing power, signals a pivotal moment for China’s AI ecosystem, potentially setting new benchmarks for integrated AI capabilities and accelerating technological independence.