DeepSeek-V4.1-Flash

What is DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is a lightweight large language model version released by DeepSeek, focused on efficient inference and fast responses.

Key features

• Supports multimodal input, natively handles images and text, and generates text autoregressively
• Uses a Causal Encoder-Decoder (CED) architecture; only 8B parameters are activated in the prefill stage and 16B in the decode stage
• Supports a context length of up to one million tokens
• Uses SWA Bounded Replay technology to reduce persistent KV cache usage to approximately 1/8 of the original
• Integrates Compressed Sparse Attention 2 (CSA2) and FP4 main KV cache, further compressing KV cache usage

Who it is for

Suitable for developers and researchers who need efficient long-context reasoning, input-intensive Agent tasks, and multimodal processing.

How to get started

1. Visit the Hugging Face page https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
2. Download the model weights and load them using the standard Hugging Face Transformers method

Compiled from public descriptions on the tool's official website / README / launch page. Pricing and user numbers have not been verified; please check the official website. Some sections were drafted with AI assistance. Written in Chinese and English; other languages are machine-translated.