Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. has launched DeepSeek-V4.1-Flash, a new mixture-of-experts model featuring 552 billion parameters. The company reports that this model outperforms its predecessor, DeepSeek-V4-Pro, in speed, cost, and overall performance metrics.
Starting September 14, API requests previously directed to V4-Pro will be rerouted to V4.1-Flash. This transition reduces output costs for users by approximately 70%, with peak pricing set at $1.20 per million output tokens compared to $3.96 for the older version. Off-peak rates are $0.15 for input and $0.60 for output.
The model utilizes a causal encoder-decoder architecture with 8 billion active parameters during prompt processing and 16 billion during generation. Engineering efforts focused on optimizing the key-value cache, reducing its footprint to 890 bytes per token, which is one-quarter of the requirement for the V4-Flash model.
DeepSeek-V4.1-Flash achieved a score of 90.6 on the Terminal-Bench 2.1 benchmark and resolved 74.2% of tasks on the DeepSWE v1.1 test. The model is now available on Hugging Face under an MIT license, as well as through the company’s web and mobile applications.
The release coincides with an Anthropic threat intelligence report identifying DeepSeek as one of seven Chinese labs involved in distillation campaigns against Claude. Anthropic claims DeepSeek conducted over 12.1 million exchanges during a 14-day period in July.
