DeepSeek V4.1 Flash Launches With Lower Prices and Native Vision

DeepSeek has unveiled V4.1 Flash, a new model that replaces its previous V4 Flash and V4 Flash Vision Experimental versions. The launch occurred on September 10, 2026, with the API name deepseek-flash. Older Flash model names will function as aliases but now direct requests to V4.1 Flash. This model boasts a redesigned architecture with approximately 748 billion total parameters, including a 552 billion backbone and 196 billion Engram parameters.

Despite its size, only around 8 billion parameters are active per token during prefill, and 16 billion during decoding, which helps to keep inference costs lower.

One of the key features of V4.1 Flash is its native multimodal capabilities, allowing the DeepSeek-ViT encoder to process images alongside text. The model supports image resolutions up to 1344 × 1344 pixels. V4.1 Flash also boasts an extended context window of up to 1 million tokens and a maximum output length of 384,000 tokens. Developers can adjust the reasoning effort from 1 to 100 to balance cost and accuracy.

The model employs a mixture-of-experts architecture with causal encoder-decoder design, compressed sparse attention, and a vision support system. DeepSeek V4.1 Flash was trained on 45 trillion multimodal tokens and uses the MIT open-source license. It has outperformed its predecessors, V4 Flash and V4 Flash Vision Experimental, in several benchmarks.

DeepSeek's own benchmarks show significant improvements over V4 Flash, with V4.1 Flash achieving higher scores in agent and coding performance tests. Independent testing by Artificial Analysis and Vals.ai also praised the model's performance, placing it at the top among open-weight models. However, V4.1 Flash does not lead every benchmark and may perform differently in specialized evaluations.