DeepSeek · 2026-08-31 · major
DeepSeek-V4-Flash-Vision-Exp weights go public — 305B multimodal MoE under MIT
DeepSeek-V4-Flash-Vision-Exp is now downloadable. DeepSeek published the 305B multimodal model's weights on Hugging Face under an MIT license, ten days after it launched as an API-only preview.
DeepSeek's first multimodal V4 model leaves API-only preview — 305B weights, MIT license, and inference code on Hugging Face.
Quick facts
| Maker | DeepSeek |
|---|---|
| Parameters | 305B total |
| License | MIT |
| Availability | Open weights + DeepSeek API |
| Status | Experimental ("Exp"), not a production release |
| Serving | SGLang with DSpark speculative decoding |
| What's new | Weights and inference code, previously API-only |
Benchmarks
| DeepSeek-V4-Flash-Vision-Exp | 36.5% | |
|---|---|---|
| DeepSeek-V4-Flash-0731 | 26.2% | |
| Claude Opus 4.8 | 39.4% |
| DeepSeek-V4-Flash-Vision-Exp | 83.9% | |
|---|---|---|
| DeepSeek-V4-Flash-0731 | 82.7% | |
| Claude Opus 4.8 | 85% |
What is it?
The weights for DeepSeek-V4-Flash-Vision-Exp are now public. DeepSeek uploaded the full 305B-parameter checkpoint to Hugging Face on 31 August under an MIT license, together with a tokenizer and a PyTorch inference implementation. The model first appeared on 21 August as an API-only preview with nothing to download.
How does it work?
A vision encoder and an aligner sit on top of the existing DeepSeek-V4-Flash text stack, which uses DFlash attention, a Mixture-of-Experts feed-forward layer, and Hyper-Connections. The published repository covers all of those parts plus the DSpark forward path, so SGLang can serve the model with DSpark speculative decoding turned on.
Why does it matter?
An MIT license lets anyone run, fine-tune, or resell DeepSeek-V4-Flash-Vision-Exp without asking DeepSeek first. Teams building agents that read screenshots, charts, and documents can now keep that work on their own hardware, which was impossible while the model existed only behind the DeepSeek API.
Who is it for?
teams self-hosting multimodal agents
Frequently asked questions
- Can I use DeepSeek-V4-Flash-Vision-Exp in a commercial product?
- Yes. DeepSeek released DeepSeek-V4-Flash-Vision-Exp under the MIT license, one of the most permissive open-source licenses. MIT allows commercial use, modification, private forks, and redistribution, and the only condition is keeping the copyright and license notice. No separate agreement with DeepSeek is needed to serve it.
- How does DeepSeek-V4-Flash-Vision-Exp compare to Claude Opus 4.8?
- DeepSeek's own model card puts DeepSeek-V4-Flash-Vision-Exp slightly behind Claude Opus 4.8 on most tasks: 36.5 versus 39.4 on ApexBench and 83.9 versus 85.0 on Terminal Bench 2.1. It leads on two benchmarks, scoring 59.3 against 58.0 on DeepSWE and 35.0 against 34.0 on ZeroBench.
- Why does DeepSeek-V4-Flash-0731 score so much lower on ApexBench?
- DeepSeek-V4-Flash-0731 cannot see images. A footnote on the model card explains that for ApexBench and Agents' Last Exam the older model simply ignores the multimodal parts of the input, so its 26.2 reflects a text-only attempt at a visual task rather than a weaker model overall.
- Does the open-weights release replace the DeepSeek API version?
- No. DeepSeek-V4-Flash-Vision-Exp stays available through the DeepSeek API at V4-Flash rates, where images are tokenized at up to 384 tokens each and a free Files API lets you upload an image once and reference it by ID. The Hugging Face weights are an additional way to run the same model.
- What does the "Exp" in DeepSeek-V4-Flash-Vision-Exp mean?
- Exp marks it as experimental. DeepSeek describes DeepSeek-V4-Flash-Vision-Exp as the first experimental multimodal model in the V4 family rather than a finished production tier, so behaviour and support can change. The stable text-only siblings remain DeepSeek-V4-Flash and the larger DeepSeek-V4-Pro.
Try it
sglang serve --model-path deepseek-ai/DeepSeek-V4-Flash-Vision-Exp --tp 4 --speculative-algorithm DSPARK