DeepSeek released the V4.1-Flash model on September 10, featuring a 1 million token context window. The model utilizes a Mixture-of-Experts architecture with 552 billion total parameters, activating 8 billion for input and 16 billion for output tasks.
The new architecture employs an asymmetric Causal Encoder-Decoder design. The KV cache requirement is reduced to approximately 890 bytes per token, a 75% decrease compared to the previous V4-Flash model, allowing for higher concurrent user capacity on existing hardware.
The model supports multimodal processing of visual and text data. Benchmarks indicate that V4.1-Flash outperforms the company's V4-Pro model and competes with GPT-5.6 and Kimi K3. DeepSeek has begun routing V4-Pro requests to the new model.
The company released the model weights under an MIT license on Hugging Face. Developers can also access the model via the DeepSeek API using the deepseek-flash endpoint. Updated pricing for API services will take effect on September 14.
DeepSeek is reportedly preparing for an IPO on the Shanghai STAR Market. The release of this model is viewed as a demonstration of technical progress ahead of the company's potential public offering.
