DeepSeek raises prices on three V4.1 Flash models

DeepSeek V4.1 Flash input price rose from $0.051 per million tokens to $0.3 per million tokens, a 488.2% increase, while output price rose from $0.567 per million tokens to $1.2 per million tokens, a 111.6% increase. The model has 763.2B parameters and a 1,048,576-token context window.

DeepSeek V4.1 Flash 1017 input price rose from $0.006 per million tokens to $0.15 per million tokens, a 2400.0% increase, while output price rose from $0.5 per million tokens to $0.6 per million tokens, a 20.0% increase. DeepSeek V4.1 Flash 1017 (batch) input price rose from $0.011 per million tokens to $0.05 per million tokens, a 354.5% increase, while output price fell from $0.2 per million tokens to $0.15 per million tokens, a 25.0% decrease.

deepseek-ai announced DeepSeek-V4.1-Flash. According to deepseek-ai, it is a multimodal Mixture-of-Experts model with 552B backbone parameters that supports contexts of up to one million tokens and natively processes images and text, generating text autoregressively. According to deepseek-ai, it adopts a Causal Encoder-Decoder architecture with a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder, activates only 8B parameters per token during prefill and 16B during decode, and uses 1 shared expert and 384 routed experts per MoE layer, activating 6 routed experts per token. According to deepseek-ai, Engram conditional memory has 196B parameters and the model was trained from scratch on a multimodal corpus comprising 45T tokens.

According to deepseek-ai, SWA Bounded Replay reduces the persistent KV cache footprint to roughly 1/8 of that of DeepSeek-V4-Flash, FP4 main KV caching uses E2M1 format with one E4M3 scale per 16 channels, and these designs reduce the global KV cache footprint to 890 bytes per token, roughly 1/4 of DeepSeek-V4-Flash.