Event date · · DeepSeek

DeepSeek released DeepSeek-V4-Flash-Vision-Exp, its first experimental multimodal model in the V4 family

DeepSeek 深度求索Chinese AIOpen weights
FACT STATEMENT

DeepSeek released DeepSeek-V4-Flash-Vision-Exp on Hugging Face under the MIT license. It is an image-text-to-text model built on DeepSeek-V4-Flash with added visual modules. Compared to DeepSeek-V4-Flash-0731, it improves multimodal agent benchmarks: ApexBench (Pass@1) 36.5 vs 26.2, Agents' Last Exam 27.3 vs 25.2, while text-only agent performance remains comparable (Terminal Bench 2.1 83.9 vs 82.7).

China context

Original name
深度求索
Outside China
Open weights · huggingface.co
Claims
Company-reported; not yet independently evaluated
For builders
Developers outside China can download the model weights from Hugging Face and integrate multimodal capabilities under the MIT license.
For investors
The release signals DeepSeek's continued investment in open multimodal models, which may pressure other open-weight competitors.
What happened

DeepSeek released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model based on DeepSeek-V4-Flash. The model adds visual understanding capabilities and shows improved performance on multimodal agent benchmarks while maintaining text-only agent performance. The repository includes tokenizer, prompt encoding reference, and a minimal PyTorch inference implementation, with support for vLLM and SGLang.

Technical significance

The model incorporates a vision encoder and aligner, DFlash attention, MoE, Hyper-Connections, and a DSpark forward path. It supports OpenAI-style JSON content blocks and compact <image path</image> TXT notation for prompt encoding. Inference can be run on a single 4×GB300 node with vLLM, and SGLang supports speculative decoding via DSpark.

Industry impact

DeepSeek is extending its open-weights model family into multimodal capabilities, directly competing with other open multimodal models. The MIT license allows broad commercial use, which may accelerate adoption among developers and enterprises.

Decision value

The MIT license and availability on Hugging Face lower barriers for developers to integrate multimodal capabilities into applications. The model's improved multimodal agent performance could be valuable for tasks involving visual inputs, such as UI automation or document understanding.

What to watch

Observable next signals include whether DeepSeek releases a non-experimental multimodal model, whether the model is integrated into DeepSeek's API or chat platform, and how the community benchmarks it against other open multimodal models.

CHINA AI WEEKLY

Get the week in Chinese AI, in English.

One weekly issue of verified model, company, robotics and policy changes, each with its original source and outside-China availability.