DeepSeek Posts Downloadable Weights for a V4 Model That Reads Images

DeepSeek has published downloadable weights for DeepSeek-V4-Flash-Vision-Exp, a model that accepts images as well as text. The repository appeared on the company's Hugging Face account at 06:16 UTC on Aug. 31, 2026, according to the platform's own metadata.

The model card calls it "our first experimental multimodal model in the DeepSeek-V4 family," built on the DeepSeek-V4-Flash architecture "by incorporating visual modules and undergoing continued training to unlock visual understanding capabilities." The "-Exp" in the name is DeepSeek's own marker for experimental, and the company does not present it as a general release.

The weights carry an MIT license. The repository's LICENSE file is the standard MIT text, permitting use, copying, modification, redistribution and sale without restriction, which is broader than the custom licenses that many open-weight releases attach to commercial use. Shipped with the weights are the tokenizer, a prompt-encoding reference and a minimal PyTorch inference implementation. The repository does not include training code or training data, so the release is open weights rather than open source.

DeepSeek's own comparison table on the model card reports gains over its earlier text-only DeepSeek-V4-Flash-0731 on tasks that mix pictures and text: 36.5 against 26.2 on a benchmark the card names ApexBench, and 27.3 against 25.2 on one it calls Agents' Last Exam. The card notes that the older model "ignores the multimodal elements in the input" on both. On text-only tasks it reports 83.9 against 82.7 on Terminal Bench 2.1.

Those figures are DeepSeek's own, run with the company's evaluation harness at settings the card lists. Hugging Face marks the three results it indexes as unverified, and no independent evaluation of the model has been published since the weights went up.

DeepSeek's instructions include a vLLM command that it says serves the model on a single node of four GB300 GPUs.