it is coming.
Leaked specs suggest a massive 671B parameter model with 37B active parameters per token.
A viral leak from a Chinese AI researcher on X (formerly Twitter) has sparked intense speculation about DeepSeek's next-generation model. The post, titled 'it is coming,' hints at DeepSeek-V3, rumored to be a colossal 671 billion parameter model. Crucially, it is suggested to utilize a Mixture of Experts (MoE) architecture, a design that routes each input through only a subset of the total neural network. This specific leak claims the model would activate approximately 37 billion parameters per token, a fraction of its total size, which is key for efficiency.
If accurate, this architecture would allow DeepSeek-V3 to potentially rival the performance of giants like OpenAI's GPT-4 while being significantly more cost-effective to run. The MoE approach is a cutting-edge strategy to scale model capability without a linear increase in computational demand for inference. The leak has not been officially confirmed by DeepSeek, but it aligns with the industry's competitive push towards larger, more efficient models. It signals DeepSeek's ambition to be a top-tier contender in the global foundation model race, moving beyond its strong regional presence in China.
- Leaked specs point to a 671B parameter model dubbed DeepSeek-V3.
- Uses a Mixture of Experts (MoE) design, activating only ~37B parameters per token for efficiency.
- If real, it represents a major competitive move against models like GPT-4 in capability and cost.
Why It Matters
A confirmed model of this scale would intensify the global AI arms race, offering powerful, cost-efficient alternatives for developers and enterprises.