DeepSeek V4 Trillion-Parameter Model Set for End-of-April Release Amid First Funding Round

Chinese AI company DeepSeek plans to release its trillion-parameter flagship model DeepSeek V4 by end of April, featuring 1.25-1.6 trillion total parameters with a Mega MoE architecture that activates thousands of experts while maintaining 37 billion active parameters per inference. The release coincides with DeepSeek's first external fundraising round seeking at least $300 million at over $10 billion valuation, marking a strategic shift from its historically independent operation model backed by hedge fund High-Flyer Quant. DeepSeek V4 will be China's first trillion-parameter model fully independent from NVIDIA hardware, running entirely on domestic chips like Huawei Ascend, a move NVIDIA CEO Jensen Huang called potentially terrible for U.S. interests. The fundraising aims to address talent competition after several core researchers departed to competitors like Xiaomi and ByteDance, where rivals reportedly offer compensation packages approaching 100 million yuan annually, while DeepSeek has fallen behind in monthly active users and faces a 15-month gap since its last major update.

DeepSeek V4's Trillion-Parameter Architecture and Technical Capabilities

• ▪DeepSeek V4 supports a context window of up to one million tokens.
• ▪DeepSeek engineers transitioned from NVIDIA's CUDA ecosystem to Huawei's CANN architecture to accomplish the underlying migration for V4.
• ▪DeepSeek V4 will be fully compatible with domestic AI chips like Huawei's Ascend during the inference phase.
• ▪DeepSeek V4's inference speed is 35 times faster than V3's according to internal testing.
• ▪DeepSeek V4's Mega MoE architecture increases the number of experts activated during a single forward inference from V3's 256 to the order of thousands.
• ▪DeepSeek V4 is China's first trillion-parameter large model to completely break free from reliance on NVIDIA hardware across the entire compute pipeline.
• ▪DeepSeek V4's total parameter count would hit 1.25 trillion with a 48-layer structure.
• ▪DeepSeek V4's parameter count represents a doubling or greater increase compared to the previous V3 model's 670 billion parameters.
• ▪DeepSeek is expected to officially release its new flagship large language model, DeepSeek V4, by the end of April.
• ▪DeepSeek V4's single-layer MoE structure has approximately 25.37 billion parameters.
• ▪DeepSeek V4 maintains the number of parameters activated per inference around 37 billion.
• ▪DeepSeek V4's total parameter count would reach about 1.6 trillion with a 60-layer design.
• ▪DeepSeek V4's energy consumption is reduced by 40% compared to V3.

First External Fundraising Round and Strategic Shift

• ▪Market rumors suggest that competitors have offered annual compensation packages approaching 100 million yuan to top researchers at DeepSeek.
• ▪DeepSeek is planning its first external fundraising round with a valuation of no less than $10 billion.
• ▪DeepSeek aims to secure at least $300 million in its first external fundraising round.
• ▪High-Flyer Quant is estimated to have brought Liang Wenfeng over $700 million in revenue in 2025 alone.

Competitive Pressures and Ecosystem Challenges

• ▪ByteDance's Doubao monthly active users across all channels exceeded 331 million as of March 2026.

Perspective of NVIDIA and Jensen Huang

• ▪Jensen Huang characterizes AI models optimized for non-NVIDIA domestic hardware platforms as potentially terrible outcomes for United States interests.

Perspective of Chinese AI competitors

• ▪Chinese AI competitors are offering DeepSeek's top researchers annual compensation packages approaching 100 million yuan to poach talent.

1 source