DeepSeek V4.1 Flash API: 1M context, image input, reasoning control
DeepSeek V4.1 Flash (deepseek-v4.1-flash) is available on FastMetal through the API and in chat. Two things set it apart from the older deepseek-v4-flash: it reads images, and its reasoning goes down to low or off entirely.
DeepSeek describes it as a sparse mixture-of-experts model with 8B parameters active per token, and the company's first model built on its CED (Causal Encoder-Decoder) architecture.
Capabilities
- 1M-token context: 1,048,576 tokens in, up to 384,000 tokens out.
- Image input: send images alongside text. This is a first for DeepSeek's Flash line;
deepseek-v4-flashis text only. - Graded reasoning:
low,highormax, or off.deepseek-v4-flashonly offershighandxhigh, so there is no light setting. - Tool calling and structured output: function calling, and
json_schemaresponses withstrict.
Set reasoning explicitly
Reasoning tokens are billed as output tokens. If you leave reasoning at its default and ask a short question, the visible answer can be a few dozen words while the output token count runs into the thousands.
When we added the model in September 2026, reasoning was on by default at high. We ran the 43 prompts from our comparison set through it. One unremarkable request (rework a recipe with different ingredients) used 7,865 completion tokens, and 6,389 of them were thinking. The answer took 381 seconds. About 80% of the output tokens on that request went to reasoning the reader never sees. The model has since moved to a different host, and the default can change with the host, so do not rely on it.
Pick the level for the job:
- Classification, extraction, reformatting:
{"enabled": false} - Everyday code generation and Q&A:
{"effort": "low"} - Hard design or math problems:
{"effort": "high"}or"max"
In FastMetal chat, the thinking control sits next to the model picker. For the API, see reasoning control.
Send the reasoning object shown below. The gateway translates it for whichever host serves the model, while a bare OpenAI-style reasoning_effort field can be ignored.
Calling it
The API is OpenAI-compatible. Set model to deepseek-v4.1-flash:
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4.1-flash",
"messages": [{"role": "user", "content": "Explain the cause of this error in one sentence."}],
"reasoning": {"effort": "low"}
}'
For an image, use the standard OpenAI image_url content part:
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4.1-flash",
"messages": [{
"role": "user",
"content": [
{"type": "text", "text": "Read the error message on this screen."},
{"type": "image_url", "image_url": {"url": "https://example.com/screenshot.png"}}
]
}],
"reasoning": {"effort": "low"}
}'
V4 Flash or V4.1 Flash
deepseek-v4-flash | deepseek-v4.1-flash | |
|---|---|---|
| Context | 1M tokens | 1M tokens |
| Image input | No | Yes |
| Reasoning levels | high / xhigh | low / high / max, or off |
Per token, deepseek-v4-flash is cheaper, and it is still a strong pick for high-volume routine work. V4.1 Flash earns its place when you need to read images, or when you want to run with reasoning at low or off. A higher per-token rate can still cost less per request if the model thinks less. Run your own prompts through both to find out.
Both are billed per million input and output tokens. Current rates are on the DeepSeek V4 Flash and DeepSeek V4.1 Flash model pages and on pricing. For a side-by-side on the same prompts, see DeepSeek V4.1 Flash vs GLM-5.3 Flash, and each model's real answers are on the comparison page.