DeepSeek V4.1 Flash

About this model

DeepSeek V4.1 Flash is a 552B-parameter multimodal Mixture-of-Experts (MoE) chat model from DeepSeek AI, activating only 8B parameters during prefill and 16B during decode thanks to its novel Causal Encoder-Decoder (CED) architecture with Compressed Sparse Attention 2 (CSA2). The model natively processes images and text, supports a 1M-token context window, and features continuously controllable reasoning effort (1–100) for trading inference cost against accuracy. It achieves 90.9 on GPQA Diamond, 90.6 on Terminal-Bench 2.1, and 74.2 on DeepSWE v1.1, demonstrating frontier-level agentic and coding performance. The model is released under the MIT License.

Technical specifications

Capabilities

Input modalities

Output modalities

Reasoning

Hybrid Default on

Knowledge horizon