The Hangzhou-based startup DeepSeek released its V4.1 Flash AI model on September 10. The system features 552 billion parameters and utilizes a mixture-of-experts architecture to reduce computational overhead by activating only specific parameters for tasks.
The model supports native multimodal capabilities, including text and image processing. It is optimized for coding and agentic workflows, which involve autonomous multi-step planning and execution.
Market reaction was immediate, with shares of competitors MiniMax and Z.ai falling by more than 8%, while Alibaba stock declined by over 2%. Investors are concerned about potential margin compression across the Chinese AI sector.
DeepSeek previously priced models at $0.14 per million input tokens and $0.28 per output token. Earlier versions were reported to be over 100 times cheaper to run than alternatives, with tests averaging $0.03 compared to $3.15 for Claude Fable 5.
The company previously trained its V3 model using 2,000 Nvidia H800 GPUs for a reported cost of $5.6 million. DeepSeek is currently preparing for an IPO on the Shanghai STAR Market to secure capital and establish a valuation benchmark.
