DeepSeek has announced the release of DeepSeek V4.1 Flash, a new model featuring a multimodal Mixture-of-Experts (MoE) architecture. According to official statements from the company’s WeChat account, the model is designed to provide high-speed inference and increased throughput while maintaining a lower operational cost compared to previous iterations.
Architectural Specifications and Performance
The V4.1 Flash model utilizes a brand-new Causal-Encoder-Decoder (CED) architecture. While the model is described as having a 552B-parameter backbone, technical analysis of the model weights on Hugging Face suggests a more complex composition. Independent observations indicate the main model consists of approximately 551.566B parameters across 40 layers, with FFN experts totaling 543.582B parameters. When including additional components such as the engram (approximately 196.929B), DSpark/MTP (14.225B), and the vision encoder (0.485B), the total parameter count is estimated by some users to reach approximately 748B.
The asymmetric architecture is designed to optimize resource allocation: only 8B parameters are activated during the input phase, while 16B are activated during the output phase. This design aims to achieve high intelligence at a reduced computational cost. DeepSeek reports that the model’s intelligence level surpasses that of its predecessor, DeepSeek V4 Pro, in various benchmark tests.
Memory and KV Cache Optimization
A significant technical focus of the V4.1 Flash release is the reduction of KV cache requirements. DeepSeek claims that compared to the previous generation, HBM (High Bandwidth Memory) requirements have dropped to one-fourth, while SSD requirements have decreased to one-eighth. This reduction is intended to lower costs for agent-style tasks where cache-hit charges are a significant factor.
The model also introduces an “engram” component. While some users on community forums like Reddit have debated the exact size and placement of these components, discussions suggest that the engram can be stored on an SSD to mitigate high VRAM requirements, albeit with a potential speed reduction of 20-30%.
API Availability and Pricing
DeepSeek has integrated V4.1 Flash into its API with native multimodal support. Users can access the model via the deepseek-flash identifier. The company has announced plans to phase out the V4 Pro model; after 12:00 Beijing time on September 14, 2026, requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at the Flash model’s unit price.
New pricing structures for V4.1 Flash take effect on September 10, 2026. The company has implemented a peak/off-peak pricing model, where off-peak rates are set at half the peak rate to encourage efficient task scheduling.
Deployment and Partnerships
Tencent (via WorkBuddy and CodeBuddy) and OpenCode have reportedly integrated DeepSeek V4.1 Flash into their services. While the model weights have been released on Hugging Face to support the open-source community, DeepSeek noted that large-scale deployment requires significant hardware resources, specifically citing the need for a 2k-GPU cluster and dedicated storage.