DeepSeek opened its eyes today. On 2026-08-21, the experimental multimodal model DeepSeek-V4-Flash-Vision-Exp went live on the DeepSeek API platform — set model='deepseek-v4-flash-vision-exp' and it works (source). The headline is the price: identical to V4-Flash, not a cent more.
What This Model Does
In short: a vision module layered onto V4-Flash's text capability, at the same price. DeepSeek says pure-text capability (agents, reasoning, world knowledge) matches the V4-Flash GA release, while vision-agent benchmarks jump sharply, bringing multimodal agent performance close to Opus-4.8 (source). For developers, DeepSeek's biggest pain point was "can't see images" — especially inside DeepSeek Harness, where uploading a screenshot returned MODEL_DOES_NOT_SUPPORT_IMAGES. That gap is now filled.
It handles JPEG, PNG, GIF, and WebP — formats are detected by actual file content, not the filename or declared MIME type — and can describe images, read text in screenshots, and analyze charts (source).
How to Pass Images
Three ways to send images: base64 inline, external URL, or Files API, all in the standard OpenAI-compatible format. Base64 is simplest for local files, an external URL lets the model fetch a publicly accessible link itself, and the Files API lets you upload once and reference a file_id repeatedly (source). Each has its own limits:
| Method | Limits | Best for |
|---|---|---|
| Base64 inline | Counts toward the 48 MiB request body limit; 32 MiB per image | Local files |
| External URL | URL up to 8192 chars; 32 MiB per image; download within 60s | Images with a public link |
| Files API | Up to 64 MiB per image, exempt from the 32 MiB check | Large images, reusing one image across requests |
DeepSeek also launched a free Files API the same day: upload an image once, reference it by file_id in later requests, and stop re-uploading the same image to save bandwidth (source).
How Billing Works
Images are tokenized by size and billed together with text, capped at 384 tokens each, at exactly V4-Flash rates. Before entering the model, every image is auto-scaled: images smaller than roughly 384×384 total pixels are scaled up proportionally, and larger ones are scaled down to about an 800×800 equivalent — so a 2000×2000 and a 5000×5000 image consume the same tokens, and no single image ever exceeds 384 (source).
Billing (model id = deepseek-v4-flash-vision-exp, per 1M tokens):
| Line item | Off-peak (CNY) | Peak (CNY) |
|---|---|---|
| Input (cache hit) | ¥0.05 / MTok | ¥0.10 / MTok |
| Input (cache miss) | ¥1.5 / MTok | ¥3 / MTok |
| Output | ¥4.5 / MTok | ¥9 / MTok |
| Per image | ≤ 384 tokens (counted as input) | — |
Peak hours are 9:00–12:00 and 14:00–18:00 Beijing time; everything else is off-peak at half the peak price (source). At peak input, one image caps out at roughly 384 × ¥3/1M ≈ ¥0.001 — effectively free to test with.
How Strong Is the Vision Capability
The official changelog lists the full vision benchmark set — Chartography at 64.3 and ZeroBench at 35.0 — a sharp leap over V4-Flash, with multimodal agent capability "close to Opus-4.8." (source):
| Benchmark | deepseek-v4-flash-vision-exp |
|---|---|
| Terminal Bench 2.1 | 83.9 |
| NL2Repo | 57.7 |
| DeepSWE | 59.3 |
| DSBench-Hard | 63.6 |
| AutomationBench (Public) | 25.7 |
| ApexBench (Pass@1) | 36.5 |
| Agents' Last Exam | 27.3 |
| Chartography | 64.3 |
| ZeroBench (Pass@5) | 35.0 |
Code Agent text tasks above were tested with DeepSeek Harness minimal mode, max effort, temperature=1.0, top_p=0.95.
Things to Watch
The model is still experimental. Clear some pitfalls up front (source):
- Only this model accepts images. Passing an image to
deepseek-v4-flashordeepseek-v4-proreturns a 400 error (This model does not support image); you must usedeepseek-v4-flash-vision-exp. - Images may only appear in user messages — putting one in system or assistant messages returns a 400.
- FIM completion isn't supported yet; fall back to a text model for completion tasks.
- Up to 600 images per request, with 64 MiB total (200 MiB with file_id); max edge 8192px, dropping to 4096px at 15+ images.
- An experimental build is better for prototyping than for a core production path.
Which to Pick
Go with this for vision work; stick to V4-Flash / V4-Pro for pure-text tasks. If you're doing screenshot understanding, mixed text-image document parsing, chart analysis, or vision-capable agents, use deepseek-v4-flash-vision-exp at the same cost as Flash. If it's just coding or batch processing, you don't need the vision overhead — deepseek-v4-flash is cheaper.
For a fixed monthly spend, DeepSeek has no official monthly coding plan; consider third-party bundles like Volcengine, CtCloud, and Baidu Qianfan, which meter DeepSeek against plan quota and aren't subject to the official concurrency cap or price hike.