DeepSeek-V4.1-Flash Release Notes

DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts (MoE) model designed for higher throughput, faster inference, and native visual understanding. It introduces an asymmetric architecture that allows it to outperform previous flagship models, including DeepSeek-V4-Pro, while reducing operational costs.

Asymmetric Architecture and Model Intelligence

DeepSeek-V4.1-Flash utilizes a 552B-parameter MoE structure with a new Causal Encoder-Decoder architecture. This design optimizes resource allocation by using only 8B active parameters for input processing and 16B active parameters for output generation.

Intelligence gains are driven by a combination of new pre-training methods and larger-scale Reinforcement Learning (RL) post-training. According to DeepSeek, these improvements allow V4.1-Flash to achieve benchmark results that exceed those of the DeepSeek-V4-Pro flagship model.

KV Cache Optimization and Cost Reduction

DeepSeek-V4.1-Flash significantly reduces the memory and storage requirements for its Key-Value (KV) cache compared to the previous generation. The model requires:

  • 1/4 of the High Bandwidth Memory (HBM) previously required.
  • 1/8 of the SSD storage previously required.

This compression of the KV cache is specifically targeted at reducing the costs associated with cache-hit charges, which DeepSeek notes often represent a large portion of agent-based operational costs.

API Integration and Model Lifecycle

DeepSeek-V4.1-Flash is now available via the DeepSeek API under the model identifier deepseek-flash. The release includes the following lifecycle changes:

  • Retirements: V4-Flash and V4-Flash-Vision-Exp are now retired.
  • Compatibility: Requests sent to deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1-Flash.
  • V4-Pro Transition: DeepSeek is phasing out V4-Pro. Starting September 14, 2026, at 04:00 UTC, all deepseek-v4-pro requests will be routed to V4.1-Flash at V4.1-Flash pricing until the launch of V4.1-Pro.

Official partners WorkBuddy (including CodeBuddy) and OpenCode now provide full support for V4.1-Flash.

Pricing and Deployment

DeepSeek has implemented new pricing effective September 10, 2026, at 04:00 UTC, citing the efficiency of the V4.1-Flash architecture. The pricing model continues to use peak and off-peak rates, with off-peak rates set at 50% of peak rates to balance demand.

For open-source deployment, the model and technical report are available on Hugging Face. DeepSeek is also offering consultation for large-scale deployments requiring 2,000 GPUs and a storage cluster.

Sources