Model comparison · 2026-08-23
DeepSeek V4 Flash Vision: a guide to the 1M-context model that sees images
DeepSeek V4 Flash Vision Exp explained: accepted image input formats, the one-million-token context window, peak-hour pricing mapped to local time, and tool calling support with measurement results.
What is DeepSeek V4 Flash Vision Exp?
DeepSeek V4 Flash Vision Exp is DeepSeek's experimental model that accepts image input. The Exp in the name means experimental; the model can be used in production flows, but the provider is more likely to change its behaviour or interface than on other rows.
Three properties set it apart: a 1,000,000-token context window, a 393,216-token output ceiling, and image understanding. Together they make a surface well suited to processing a long document together with its images in a single request.
- Context window: 1,000,000 tokens.
- Maximum output in one response: 393,216 tokens (384K).
- Image input: remote https URL and base64 data URI.
- Tool calling is supported; thinking is always on and cannot be turned off.
Which image formats can you send?
The model accepts images in two forms: a direct https URL, or a base64-encoded data URI. Both are passed through the OpenAI-compatible image_url field, so existing image-handling code largely stays as it is.
There is one detail worth knowing about remote URLs. This behaviour was measured on 22 August 2026 against four different hosts: an image hosted on Baidu returned 200 and the model described the content correctly, while Wikimedia, jsDelivr and alicdn each returned 400 Failed to download image.
That result does not mean the model lacks remote URL support — it has it. What the test measured is whether the host is reachable from DeepSeek's network. The practical conclusion: unless you serve images from your own server or a CDN you know to be reachable, a base64 data URI is the more reliable path, because it depends on no external fetch at all.
Sending an image as a base64 data URI: the reliable path that depends on no external fetch
import base64
import os
from openai import OpenAI
client = OpenAI(
base_url="https://llmtr.com/v1",
api_key=os.environ["LLMTR_API_KEY"],
)
with open("invoice.png", "rb") as handle:
encoded = base64.b64encode(handle.read()).decode("ascii")
response = client.chat.completions.create(
model="deepseek/deepseek-v4-flash-vision-exp",
messages=[
{
"role": "user",
"content": [
{"type": "text", "text": "What is the total amount and date on this invoice?"},
{
"type": "image_url",
"image_url": {"url": f"data:image/png;base64,{encoded}"},
},
],
}
],
max_tokens=1024,
)
print(response.choices[0].message.content)
Peak-hour pricing: when is it expensive?
DeepSeek applies time-windowed pricing. One rate card applies during standard hours and exactly double that during peak hours. The multiplier applies separately to input, cache reads and output alike.
The peak window is defined in Beijing time: weekdays 09:00-12:00 and 14:00-18:00. Because Turkey stays on UTC+3 all year, the window maps to a fixed local range that is easy to plan around.
In Turkish local time the peak hours are weekdays 04:00-07:00 and 09:00-13:00. Every other hour bills at the standard rate. In addition, from 23 August 2026 weekends are entirely standard rate — Saturday and Sunday sit on the cheap side regardless of the hour.
The practical consequence: moving batch jobs, overnight data processing and weekend queues outside those windows halves the cost without changing model. Most of the Turkish working day already falls on the standard rate; the ranges that need attention are the early morning and the late-morning block.
| Item | Standard hour | Peak hour (2x) |
|---|---|---|
| Input | USD 0.22 | USD 0.44 |
| Cache read | USD 0.007 | USD 0.014 |
| Output | USD 0.66 | USD 1.32 |
| Turkish local time | All other hours and weekends | Weekdays 04:00-07:00 and 09:00-13:00 |
Thinking is always on: why max_tokens matters
Thinking is fixed on this model. It reasons step by step and no setting turns that off; there is no user-defined reasoning_effort option. Do not look for the toggle that some other models offer.
The billing consequence is direct: even a short question spends thinking tokens, and thinking tokens bill as output. With output at USD 0.66 in standard hours and USD 1.32 at peak, the length of the reasoning chain can become the dominant cost line.
Set max_tokens deliberately as a result. Too low and the model is cut off before it finishes reasoning, leaving you with an incomplete answer. Too high and you absorb the cost of long reasoning chains. Measuring a handful of samples from your own workload and looking at the completion_tokens distribution beats guessing.
What are the 1M context and 384K output ceiling good for?
A 1,000,000-token context means you can hand the model a very large body of material in one request. Roughly 2,500 pages of text fit in that window. Long contract sets, an entire module of a codebase, or a year of meeting notes can be processed in a single call.
The 384K output ceiling is markedly higher than most models offer. For work where the output is as long as the input — translation, rewriting, or turning a whole document into structured output — that ceiling removes a real constraint.
That said, filling the context is not always the right approach. Input bills too, and a one-million-token request alone costs USD 0.22 at the standard rate. Serving repeated material from the prompt cache is more than thirty times cheaper, given cache reads at USD 0.007.
Tool calling and access through LLMTR
The model supports function calling and can be used in coding agents and flows that reach external systems. Image input and tool calling combine in the same request: you can hand the model a screenshot and ask it to call a tool based on what it sees.
Access works with OpenAI-compatible clients; set the base URL to https://llmtr.com/v1 and pass deepseek/deepseek-v4-flash-vision-exp as the model identifier. LLMTR does not mark up model prices; the platform margin applies only to credit top-ups.
Frequently asked questions
Does DeepSeek V4 Flash Vision accept images from a remote URL?
Yes, it accepts both remote https URLs and base64 data URIs. In the 22 August 2026 measurement some hosts returned 400 Failed to download image, which reflects that host's reachability from DeepSeek's network rather than a model limitation. For sources whose reachability you cannot confirm, base64 is more reliable.
What are the peak hours in Turkish local time?
Weekdays 04:00-07:00 and 09:00-13:00 are peak, where the price doubles. Every other hour, plus all of Saturday and Sunday, bills at the standard rate.
Can I turn the model's thinking off?
No. Thinking is fixed on this model and cannot be disabled; no user-defined reasoning_effort option is offered. Since thinking tokens bill as output, plan your max_tokens value accordingly.
Does it make sense to fill the one-million-token context?
Not always. Input bills too, and a one-million-token request costs USD 0.22 at the standard rate on its own. Serving repeated material from the prompt cache is far cheaper, with cache reads at USD 0.007.