by DeepSeek
Experimental vision-enabled version of DeepSeek V4 Flash with image understanding; open weights under MIT, now retired on the API.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
DeepSeek-V4-Flash-Vision-Exp is the first experimental multimodal model in the DeepSeek-V4 family. It extends the DeepSeek-V4-Flash architecture with a vision encoder and aligner, followed by continued training to add visual understanding. Output is text. The Hugging Face card reports substantially better multimodal agent results than DeepSeek-V4-Flash-0731 while keeping comparable text-only agent performance.
Vendor-reported benchmarks compare it with DeepSeek-V4-Flash-0731 and Opus-4.8. Examples: Terminal Bench 2.1 83.9 (Opus-4.8: 85.0), ApexBench Pass@1 36.5 (39.4), Agents' Last Exam 27.3 (25.7) and ZeroBench Pass@5 35.0 (34.0). These are self-reported numbers produced with DeepSeek's own agent harness. YouTube creators describe it as roughly on par with Opus 4.8 on several agentic tasks, and one cites about 305B parameters and 168 GB of weights, which the provided pages do not confirm.
The weights are on Hugging Face under the MIT license, together with a tokenizer, prompt encoding reference and a minimal PyTorch inference implementation. Through the API, images were sent in OpenAI-compatible Chat Completions or Responses format. Per the current API docs, the legacy model name deepseek-v4-flash-vision-exp is still accepted, but the model is retired and requests are served by the latest deepseek-flash model, which accepts images (JPEG, PNG, GIF, WebP).
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Vision-enabled version of DeepSeek V4 Flash, said to be roughly on par with Claude Opus 4.8 in many cases. Now open-sourced on Hugging Face at 305B parameters and about 168 GB.
Adds image, video and document understanding to V4 Flash while matching its text performance. It matches Opus 4.8 on several agentic coding benchmarks; API only for now.
Closest neighbours by meaning, not just by category.
Model · DeepSeek
Open-weights multimodal DeepSeek model in the V4 family, offered via API, chat and downloadable weights.
Model · DeepSeek
Open-weights Flash-tier model in DeepSeek's V4 family, offered via chat and a low-cost API.
Model · DeepSeek
Chinese AI lab behind the open-weight DeepSeek-V4 and DeepSeek-R1 models, with a chat app and low-cost API.
Model · DeepSeek
DeepSeek's V4 Pro large language model, available via API and as an open model, with low per-token pricing.
Model · StepFun
Open-weights multimodal flash model from StepFun for agentic tasks involving images, interfaces, web search and long runs.
Model · Ant Group
Ant Group's multimodal mixture-of-experts model (560B total, 25B active) in the Ling family, served via API.