Key Points
DeepSeek-V4-Flash-Vision-Exp launched August 21, 2026, adding multimodal support to V4-Flash.
Files API lets developers reuse uploaded images for free via file_id references.
Images cost up to 384 tokens each, billed at standard V4-Flash rates.
Vision model nears Opus-4.8 performance on multimodal agent benchmarks, DeepSeek says.
DeepSeek rolled out DeepSeek-V4-Flash-Vision-Exp on its API platform on August 21, 2026. The experimental multimodal model matches DeepSeek-V4-Flash on text tasks, including reasoning and agents. It adds vision support and a free Files API for image reuse. The launch targets developers building multimodal agent workflows at scale.
What the New Vision Model Brings
DeepSeek-V4-Flash-Vision-Exp accepts both text and image inputs in a single request. Developers activate it by setting model=’deepseek-v4-flash-vision-exp’ in their API calls. The model supports JPEG, PNG, GIF, and WebP image formats natively.
On pure-text benchmarks, the vision model performs on par with standard DeepSeek-V4-Flash. On multimodal agent benchmarks, it shows a sharp jump over its predecessor. DeepSeek says the gain brings its agent performance close to Anthropic’s Opus-4.8 on visual tasks.
Files API Cuts Repeated Upload Costs
Free Reuse via file_id
The new Files API lets developers upload an image once and reuse it across multiple requests. This works by referencing a returned file_id instead of re-sending the file. DeepSeek confirmed the Files API remains free to use for all accounts.
- Upload once, reference repeatedly through file_id
- No repeated bandwidth cost for the same image
- Works across Chat Completions, Messages, and Responses API formats
Billing and Technical Limits
Images are tokenized for billing at up to 384 tokens each, charged at standard V4-Flash rates. Requests can carry mixed text and image content in a single call. DeepSeek caps total request size at 48 MiB, including any base64-encoded image data.
How It Fits Into the V4 Lineup
DeepSeek-V4-Flash carries 284 billion total parameters, with 13 billion active per request. The model supports a 1-million-token context window across both thinking and non-thinking modes. DeepSeek-V4-Pro-0813 went GA on App, Web, and API on August 13, 2026.
The vision variant stays experimental for now and sits outside the official model catalog. DeepSeek has not confirmed a timeline for graduating it to general availability. Developers should expect API behavior to shift before a stable release.
Related AI and Chip Stocks to Watch
DeepSeek’s open-weight releases keep pressuring pricing across the AI model market. Investors tracking this space often watch chipmakers and rival AI platforms tied to the same demand cycle:
- Nvidia (NVDA): GPU demand tied to inference workloads from models like V4-Flash
- AMD (AMD): competing accelerator supplier for large-scale AI deployment
- Alibaba (BABA): parent ecosystem for competing Chinese LLM development
- Baidu (BIDU): runs its own Ernie model line competing directly with DeepSeek
Analysts’ View
DeepSeek’s vision launch signals faster iteration on multimodal agents from Chinese AI labs. Free image reuse through the Files API lowers costs for developers building at scale. The real test comes once the model exits its experimental tag.
Disclaimer:
The content shared by Meyka AI PTY LTD is for research and informational purposes only. Meyka is not a financial advisory service, and the information provided should not be treated as investment or trading advice.
What brings you to Meyka?
Pick what interests you most and we will get you started.
I'm here to read news
Find more articles like this one
I'm here to research stocks
Ask Meyka Analyst about any stock
I'm here to track my Portfolio
Get daily updates and alerts (coming March 2026)