DeepSeek releases weights for multimodal V4-Flash-Vision-Exp

DeepSeek has released the experimental multimodal model DeepSeek-V4-Flash-Vision-Exp, a 305-billion parameter (MoE) model, on Hugging Face under the MIT license, just 10 days after the release of the text-based V4-Flash.

Compiled by Sergey KostenchukPublished 2026-09-01Updated 2026-09-18

2026-09-01 Research OpenAI

🤖 DeepSeek releases weights for multimodal V4-Flash-Vision-Exp

An experimental 305-billion parameter (MoE) checkpoint has been released on Hugging Face under the MIT license — just 10 days after the release of the text-based V4-Flash. A visual encoder has been added to the V4-Flash architecture: the model can parse screenshots, read charts, and work with tools.

🌍 The first multimodal model in the open V4 lineup, focused on agents: it scored 36.5 on ApexBench compared to 26.2 for the text-based V4-Flash-0731 (which cannot see images), outperforms Claude Opus 4.8 on Agents' Last Exam and ZeroBench, but trails it on some tasks.

👤 The weights can be downloaded and run locally — quantized builds are available for llama.cpp, LM Studio, and Ollama. Without your own hardware — via the DeepSeek API (OpenAI-compatible format with image_url, up to 600 images per request).

Source 1: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp Source 2: https://api-docs.deepseek.com/guides/vision/

Sources