DeepSeek Announces V4.1 Flash Model Outperforming V4 Pro
DeepSeek announced the upcoming release of its V4.1 Flash model, introducing a new architecture and native multimodal capabilities. Internal and external testing showed that V4.1 Flash surpasses the previous V4 Pro model across all metrics, including speed, cost, and overall performance. A lighter 'Flash' tier model outperforming a top-tier 'Pro' model demonstrates rapid architectural efficiency gains in large language models, bringing down AI operational expenses for businesses. By automatically routing V4 Pro requests to V4.1 Flash at lower rates, DeepSeek immediately reduces API costs for developers. Under the new pricing scheme, off-peak input cache hits cost 0.02 RMB per million tokens, cache misses cost 1 RMB, and outputs cost 4 RMB, while peak-hour prices double. Until V4.1 Pro is released, all API traffic for V4 Pro will be automatically rerouted to V4.1 Flash and billed at the cheaper Flash rate.
## BACKGROUND
Native multimodal models integrate text, image, and other data modalities into a single unified backbone architecture during pre-training, enabling more efficient cross-modal reasoning than legacy models that chain separate visual encoders to language backbones. Additionally, modern LLM providers offer prompt caching, which stores key-value representations of repeated prompt prefixes to significantly lower compute requirements and usage costs.