DeepSeek-V4 multimodal model weights are open-sourced: visual capabilities match Opus-4.8, while text metrics show slight fluctuations
2026-09-01 08:16Models🔥 36.5 heat score
2sources
1days unfolding
36.5heat score
5mentions
SummaryAI generated
On September 1, 2026, DeepSeek made its DeepSeek-V4-Flash-Vision-Exp multimodal model weights available on the Hugging Face platform. This model is based on the V4-Flash architecture and includes a new visual module, with a parameter count of approximately 305 billion. Official data shows that this model performs similarly to V4-Flash in pure text tasks, but excels in visual capabilities such as image recognition, text reading from screenshots, and chart analysis. In some multimodal agent tasks, it approaches or surpasses Claude Opus 4.8, especially achieving similar or better results in ApexBench and Chartography tests. However, there are slight fluctuations in pure text performance metrics such as Terminal Bench and DeepSWE. The model supports formats like JPEG and PNG, and optimizes computational costs by controlling input pixels and the number of Tokens through scaling. Currently, this model has been made publicly available…
On August 31, DeepSeek made the weights of the DeepSeek-V4-Flash-Vision-Exp model available on Hugging Face. This model is based on the V4-Flash architecture and includes a new visual module, with a scale of approximately 305 billion parameters. Officially, it is said to be on par with V4-Flash in pure text tasks, capable of identifying photos, reading text from screenshots, and analyzing charts. In some multimodal agent tasks, it is close to Claude Opus 4.8. It performs similarly or better than Opus-4.8 in tests such as ApexBench and Chartography, but its pure text capability metrics such as Terminal Bench and DeepSWE show slight fluctuations. The model supports formats like JPEG and PNG; after scaling, the total pixels of the images are approximately 800×800. Each image can occupy up to 384 Tokens to control computational costs.
Meta has released the first multimodal model in its V4 family, making the model weights and minimum inference code available simultaneously. This model is designed to handle multiple modalities such as text, images, and videos uniformly. It uses the Transformer architecture and incorporates a hybrid attention mechanism to improve efficiency. The project team has provided complete training scripts and deployment guidelines, enabling it to be run on mainstream frameworks. Currently, this model is available on the Hugging Face platform, and developers can download it immediately for local testing or fine-tuning.