DeepSeek has launched V4.1 Flash, a multimodal mixture-of-experts (MoE) model designed to improve reasoning performance, inference speed and throughput while reducing deployment and operating costs. The company said the model has surpassed several leading flagship models, including its own DeepSeek V4 Pro, on selected benchmarks.
According to DeepSeek, V4.1 Flash is the smallest model in its new architecture series and features native multimodal visual understanding. It uses a new Causal-Encoder-Decoder architecture with an asymmetric input-output design, activating 8 billion parameters for input processing and 16 billion for output generation, despite having 552 billion parameters in total.
The architecture is designed to deliver greater intelligence at a lower computational cost, while supporting further scaling to larger models. DeepSeek also said the model combines a new pre-training approach with reinforcement learning post-training at a larger scale.
KV cache reduction targets AI operating costs
A key feature of V4.1 Flash is its reduced key-value (KV) cache footprint, which can help lower memory requirements and inference costs, particularly in agentic AI applications that process long contexts and repeatedly access stored information.
Compared with its previous-generation model, DeepSeek said V4.1 Flash reduces high-bandwidth memory (HBM) requirements to one-quarter and solid-state drive (SSD) requirements to one-eighth. The company added that its KV cache has become 437 times smaller than that of its first-generation model.
The reduction is expected to benefit AI agents, where cached-context processing can account for a significant share of usage costs.
API updates and pricing changes
DeepSeek V4.1 Flash is available through the DeepSeek API under the model name deepseek-flash, with native multimodal capabilities. The previous V4 Flash and V4 Flash Vision Exp models have been discontinued, with their existing API names temporarily routed to V4.1 Flash for compatibility.
DeepSeek also announced plans to phase out V4 Pro. From 12 noon Beijing time on September 14, 2026, until the launch of a future V4.1 Pro model, requests to deepseek-v4-pro will be routed to V4.1 Flash and charged at the latter's pricing.
The company revised V4.1 Flash API pricing from September 10, 2026, retaining a peak and off-peak pricing structure under which off-peak rates are half the peak rates. It said the revised pricing reflects improvements in model architecture and lower operating costs.
WorkBuddy, including CodeBuddy, and OpenCode have integrated V4.1 Flash as official partners, according to DeepSeek.
Open-source deployment support
DeepSeek said it would support the open-source community in adapting inference systems for the new model and explore ways to expand deployment availability. The company has invited organisations with large-scale deployment requirements and sufficient infrastructure, including GPU clusters with approximately 2,000 cards and storage clusters, to contact it.
The launch highlights DeepSeek's focus on improving the economics of large-scale AI inference while expanding the capabilities of models designed for multimodal understanding and agentic workloads.
