DeepSeek has released open weights for its first V4 vision model, giving developers a new way to run its multimodal system outside the company’s hosted API.
The Chinese AI company published V4-Flash-Vision-Exp on August 31, 2026. The model carries about 305 billion parameters and comes under an MIT license. DeepSeek had first made the model available through its API ten days earlier.
That release closes an important gap inside the V4 family. Earlier V4 models could work with text and code, but they could not process images.
The new model adds vision without replacing the existing text backbone. DeepSeek kept the 284-billion-parameter sparse Mixture-of-Experts transformer used by V4-Flash-0731. Around 13 billion parameters are active during each forward pass.
Images are handled through a separate 32-layer vision encoder. It breaks images into 14-by-14-pixel patches before visual information moves into the language model. The added vision components raise the full model size to roughly 305 billion parameters.
The MIT license may matter as much as the model itself. It allows developers to self-host, modify, fine-tune and commercially deploy the weights. Derivative versions can also be redistributed under the license terms described in the release.
DeepSeek’s benchmark numbers require more caution.
The company tested V4-Flash-Vision-Exp against its earlier V4-Flash-0731 model and Opus-4.8 using its own Harness Minimal Mode evaluation setup. That harness has not been publicly released, so independent researchers cannot yet reproduce the results using the same configuration.
On DeepSeek’s reported tests, the new model beat Opus-4.8 on three of eleven benchmarks. It scored 59.3 against 58.0 on DeepSWE, 27.3 against 25.7 on Agents’ Last Exam and 35.0 against 34.0 on ZeroBench Pass@5.
The picture changed on longer tasks. Opus-4.8 scored 69.7 on NL2Repo against DeepSeek’s 57.7. On DSBench-Hard, Opus-4.8 reached 71.7 while the DeepSeek model scored 63.6.
DeepSeek also reported that adding image capability did not hurt text performance. Its vision model finished ahead of V4-Flash-0731 on six of seven text benchmarks included in the company’s table.
The hardware requirements remain substantial.
V4-Flash-Vision-Exp ships entirely in FP8 rather than using the FP4 expert weights found in the text-only V4-Flash-0731. DeepSeek provides an official deployment example using four GB300 GPUs with vLLM. Community 4-bit versions begin around 155 GB, according to the material cited with the release.
That makes self-hosting a serious infrastructure decision rather than a normal desktop deployment.
The model accepts JPEG, PNG, GIF and WebP images. The API also works with OpenAI Chat Completions and Anthropic Messages formats. Its model identifier is deepseek-v4-flash-vision-exp.
DeepSeek has not announced a production-stable successor. The “Exp” label marks this checkpoint as an experimental release still under evaluation. Independent benchmark results also remain unavailable for now.
For developers, the open weights make the model easier to inspect and deploy. The harder question is whether independent testing will confirm the performance DeepSeek currently reports.
Follow iNews Zoombangla On Google
Open the Google follow page and tap the checkmark option to receive more updates from iNews Zoombangla in your Google news feed.
