When DeepSeek released the V4 family on 24 April 2026, it shipped two models — V4-Pro and V4-Flash — and both were text only. That changed on 21 August 2026, when DeepSeek quietly put out DeepSeek V4-Flash-Vision-Exp: an experimental checkpoint that gives the cheap, fast Flash model eyes. Then on 10 September 2026, V4.1 Flash arrived with vision trained in natively from the start of pre-training. This guide explains what the vision variant is, how it compares, what it costs, and exactly how to send images to it.
What DeepSeek V4 Flash Vision is
DeepSeek V4-Flash-Vision-Exp is a vision-capable build of DeepSeek V4 Flash, the company's efficiency-optimised flagship. Flash is a mixture-of-experts model: 284 billion total parameters, but only about 13 billion active per token. That sparse activation is what keeps inference cheap while retaining a huge knowledge base.
The "-Exp" suffix matters. This was an experimental checkpoint, not the final word. DeepSeek shipped it to put image understanding on the Flash line quickly, and then made vision native in the next generation, V4.1 Flash. V4-Pro, the larger sibling, still does not accept images at all — and no DeepSeek model currently accepts audio or video input.
Specs at a glance
- Architecture: Mixture of Experts, 284B total parameters, ~13B active per token
- Context window: 1 million tokens
- Input: text plus images (on the Vision-Exp checkpoint and V4.1 Flash)
- Licence: MIT, with open weights on Hugging Face under the deepseek-ai organisation
- API availability: DeepSeek's own API and third-party providers such as OpenRouter (listed as deepseek/deepseek-v4-flash-vision-exp)
The MIT licence is genuinely significant: you can download, run, modify, and commercialise the weights without per-token fees. Few frontier-class vision models offer that.
Benchmarks: what the numbers say
Independent model-index listings score the vision checkpoint at an Intelligence Index of 34.8 and a Coding Index of 65, with Terminal-Bench at 74.2, long-context reasoning at 81.3, GPQA Diamond at 91.3, and SciCode at 49.7. DeepSeek's own reported numbers for the base V4-Flash-0731 build include Terminal-Bench 2.1 at 82.7%, DeepSWE at 54.4%, Toolathlon-Verified at 70.3%, and CyberGym at 76.7%.
The takeaway pattern: the Flash line punches hardest in agentic and tool-use benchmarks, trading raw knowledge depth for speed, price, and function-calling reliability. The vision variant keeps that character while adding image understanding.
Pricing: why the Flash line gets attention
Third-party listings put the vision checkpoint at roughly $0.15 per million input tokens, $0.60 per million output tokens, and $0.003 per million on cache reads, with the full 1M context carrying no price premium. DeepSeek's own first-party API uses dynamic peak and off-peak pricing since 16 August 2026, so check the official pricing page for current numbers before budgeting.
Cache pricing deserves attention for vision workloads. Multimodal agent loops repeatedly send the same images — screenshots, documents, dashboards — alongside new questions. A cache-hit rate near $0.003/M means repeated queries against the same visual context cost almost nothing, which changes the economics of agents that "look" at things constantly.
How to use the vision endpoint
The easiest route is the API. DeepSeek's endpoint supports the OpenAI Chat Completions format, so you send images the same way you would to GPT-style models: as image_url content parts in the user message, either as a public URL or a base64 data URI.
On OpenRouter, the model is listed as deepseek/deepseek-v4-flash-vision-exp. A request looks like this:
POST https://openrouter.ai/api/v1/chat/completions
{
"model": "deepseek/deepseek-v4-flash-vision-exp",
"messages": [
{
"role": "user",
"content": [
{"type": "text", "text": "Extract the quarterly figures from this chart as a table."},
{"type": "image_url", "image_url": {"url": "https://example.com/q3-chart.png"}}
]
}
]
}
On DeepSeek's own API at api.deepseek.com, use the current flash model ID shown in the official docs — model IDs on this line have rotated (deepseek-v4-flash routed to V4.1 Flash from 10 September 2026), so verify the ID rather than hard-coding an old one. DeepSeek's API also supports an Anthropic-compatible format if your stack speaks that dialect instead.
Self-hosting: the open-weights option
If you want zero per-token cost and full data control, the weights are yours to download. The official checkpoints live under the verified deepseek-ai organisation on Hugging Face (ModelScope is the listed alternate mirror). Pull them with resumable downloads:
pip install huggingface_hub hf_transfer
export HF_HUB_ENABLE_HF_TRANSFER=1
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash --local-dir ./deepseek-v4-flash
Budget for a large download — at FP8 precision the weights run roughly one byte per parameter, so expect 250 to 300GB. Serving a 284B MoE needs serious GPU memory; vLLM supports the V4 Flash models, and DeepSeek notes the 1M-token context needs substantial KV cache on top of the weights. This is a datacentre or multi-GPU-workstation job, not a laptop project. Most developers are better served by the API.
Vision-Exp vs V4.1 Flash: which to use
The short version: Vision-Exp was the experiment, V4.1 Flash is the product. V4.1 Flash (released 10 September 2026) is a new generation — 552B parameters with a new causal encoder–decoder architecture — and vision was trained in from the start of pre-training rather than bolted on. If you are starting a new project today, target V4.1 Flash. The Vision-Exp checkpoint remains useful if you pinned it during testing or your provider still serves it, but expect the ecosystem to consolidate on V4.1.
What vision unlocks in practice
Image input turns the cheap Flash line into a multimodal workhorse. Practical patterns people use it for:
- Document QA at scale. Feed scanned invoices, contracts, or reports and ask targeted questions. The 1M-token context means you can pair long documents with their images in one request.
- Screenshot debugging. Paste an error screenshot or a broken UI render and ask what is wrong. Coding Index 65 with vision makes this genuinely useful, not a gimmick.
- Chart and dashboard extraction. Point it at a chart image and ask for the underlying data as structured output — the pattern shown in the API example above.
- UI-to-code. A mockup image plus "build this in Tailwind" is a realistic workflow given Flash's front-end code strength.
- Multimodal agents. Cheap cache reads make it economical for an agent to keep a screenshot in context and re-examine it across many tool-call turns.
Limitations to know before you commit
Be honest about the edges. Vision-Exp is an experimental checkpoint, so behaviour can shift between builds — pin the exact model version in production. Image understanding on a 13B-active model will not match dedicated frontier vision models on the hardest visual reasoning tasks. There is no audio or video input anywhere in the V4 line, so multimodal beyond text-plus-image is not on the table. And if you self-host, the weight download and GPU bill are the real cost — "open weights" is free as in licence, not free as in hardware.