DeepSeek V4.1 Flash: The Native Multimodal Model That's Breaking Speed Records

On September 8, 2026, DeepSeek quietly released a beta version of V4.1 Flash, an intermediate model capable of processing both text and images simultaneously. The beta is only available for 48 hours, from September 8 to September 10, providing developers limited time to evaluate its performance. DeepSeek's first native multimodal model, V4.1 Flash sets itself apart by achieving a remarkable speed of 420 tokens per second in long-text reasoning tasks, while maintaining high accuracy.

This speed is achieved through a new architecture featuring an improved attention mechanism, optimized routing in the MoE architecture, and enhanced multimodal fusion. Early tests showed the model excelling in tasks like SVG code generation, complex SQL query generation, and asynchronous architecture refactoring. In real-world applications, V4.1 Flash demonstrated impressive accuracy in identifying patterns from images, correctly recognizing a striped pattern in a photo of a person in a striped suit.

The model's architecture is a significant departure from previous versions, integrating text and image processing from the ground up, leading to unified representation, better cross-modal reasoning, and lower latency. DeepSeek's rapid release of models is part of an intensified AI race, with multiple Chinese AI models launched in recent months.

The company has also announced hiring 150 senior engineers to handle the increased complexity that comes with the new model's capabilities. Despite its advanced features, V4.1 Flash retains the same pricing as its predecessor, V4 Flash, a strategic move to compete with other models and gain market share.