DeepSeek-V4.1-Flash Model Introduction
DeepSeek-V4.1-Flash is a lightweight flagship model in DeepSeek's new architecture series. Built on a 552B-parameter MoE, it delivers intelligence beyond its class, outperforming mainstream flagship models including DeepSeek-V4-Pro on multiple benchmarks, and natively supports multimodal visual understanding.
- DeepSeek-V4.1-Flash Billing Instructions: Check peak-hour and off-peak-hour pricing.
Important
The AI-TokenHub DeepSeek-V4.1-Flash model has been specifically improved in terms of feature completeness, user experience and compatibility. This article compares the capabilities and API parameters of the official DeepSeek platform and this platform across supported scenarios, helping you quickly understand how this platform supports each scenario.
Supported Scenarios Overview
Enhanced Model Capabilities
| Scenario | Official DeepSeek Platform | AI-TokenHub |
|---|---|---|
| Image understanding | Single image no larger than 32MB | Single image no larger than 50MB, supports detail modes (low/high/auto/xhigh).See Image Understanding. |
| Video understanding | Not supported | Supported Single video file no larger than 50MB, supports the Files Id method.The frame sampling strategy is consistent with doubao-seed-2.0 and later models. See Video Understanding. |
| Structured output: json_schema | Not supported | Supported |
| Implicit cache | Supported | Supported Minimum cached tokens: 256 |
| Explicit cache | Not supported | Supported |
| Prefix continuation | Supported Supported in the beta API. | Supported • Chat API: Aligned with the official DeepSeek platform. • Responses API: Supported using this platform's existing API fields. See Prefix Continuation. |
Key API Parameters
| Scenario | Official DeepSeek Platform | AI-TokenHub |
|---|---|---|
| Maximum output length: max_tokens | Controls the length of the model answer and the model reasoning content; an error is returned when the length is exceeded. | Controls the length of the model answer and the model reasoning content; no error is returned when the length is exceeded. |
Chat API defaults to 8K in non-thinking mode and 64K in thinking mode (128K when reasoning_effort is max). | Chat API defaults to 128K. The Responses API uses max_output_tokens, defaulting to 128K. See Maximum Output Length. Note • deepseek-v4-1-flash-260910: max_tokens controls the length of the model answer and the model reasoning content, with a default value of 128k. If the reasoning content exceeds max_tokens, thinking will be truncated.• Other models: max_tokens controls only the length of the model answer and does not include reasoning content, with a default value of 4k. | |
| Reasoning length adjustment: reasoning_effort / reasoning.effort / output_config.effort | Supported by all three APIs: Chat API / Responses API / Messages API. The default value is high. | Supported by all three APIs: Chat API / Responses API / Messages API. The default value is high. |
• OpenAI format / Anthropic format: supports the levels low/high/max• Responses API format: supports the levels none/low/high/max | Supports the levels none/low/high/maxSee Reasoning Length Adjustment. | |
| Nucleus sampling threshold: top_p | • Thinking mode: the parameter takes effect. When the passed value is < 0.95, the system automatically raises it to 0.95; when the passed value is ≥ 0.95, the actual passed value is used.• Non-thinking mode: the field value is fixed at 1.0, and the passed value is ignored. | When using deepseek-v4-1-flash-260910, this parameter is aligned with the official DeepSeek platform. |
| Log probabilities: logprobs / top_logprobs | Supported | Supported To ensure consistency with the OpenAI protocol, after logprobs is enabled on this platform, the API returns the per-token log probabilities of the model's complete generated sequence in choices[].logprobs.content, including both reasoning content and the final answer. |
| Stop words: stop | Supports up to 16 strings. | • Chat API default: supports up to 4 stop words. • Chat API deepseek-v4-1-flash-260910 model: supports up to 16 strings. |
| Whether role supports developer | Not supported | Supported |
| tool_choice | In thinking mode, tool_choice set to required is not supported and the API returns an error. | In thinking mode, tool_choice set to required is supported. |
| Overall Responses API compatibility | Weak validation: unsupported fields are not validated and no errors are returned. | Strict validation: unsupported fields are still validated and an error is returned. Note If the User Agent contains codex, weak validation is also applied, but store is not supported. |
Model Usage Notes
Prefix Continuation
Prefix continuation behavior for the DeepSeek-V4.1-Flash model is as follows:
| API | Chat API | Responses API |
|---|---|---|
| Conditions to trigger prefix continuation | • The role of the last message in the messages list is assistant• Set the prefix parameter of the last message to true | • Set the role of the last message in the input list to assistant • Set partial to true to enable continuation mode (if partial is not set, no error is returned, but continuation is not enabled.) |
Image Understanding
The detail modes and scaling rules supported by the DeepSeek-V4.1-Flash model for image understanding are as follows:
The image pixel ranges in the table refer to the pixel ranges after the input image is processed according to the detail mode.
| detail mode | Image pixel range | Chat API | Responses API | Messages API |
|---|---|---|---|---|
| low | 262144 | The image is automatically scaled while maintaining its aspect ratio, and the total pixels after scaling are equivalent to a 512 × 512 image | The image is automatically scaled while maintaining its aspect ratio, and the total pixels after scaling are equivalent to a 512 × 512 image | • Images with a total pixel count smaller than 544 × 544 are enlarged while maintaining the aspect ratio, and the total pixels after enlargement are equivalent to a 544 × 544 image. • Images with a total pixel count larger than 1300 × 1300 are shrunk while maintaining the aspect ratio, and the total pixels after shrinking are equivalent to a 1300 × 1300 image. |
| high | [295936, 1690000] | • Images with a total pixel count smaller than 544 × 544 are enlarged while maintaining the aspect ratio, and the total pixels after enlargement are equivalent to a 544 × 544 image. • Images with a total pixel count larger than 1300 × 1300 are shrunk while maintaining the aspect ratio, and the total pixels after shrinking are equivalent to a 1300 × 1300 image. | • Images with a total pixel count smaller than 544 × 544 are enlarged while maintaining the aspect ratio, and the total pixels after enlargement are equivalent to a 544 × 544 image. • Images with a total pixel count larger than 1300 × 1300 are shrunk while maintaining the aspect ratio, and the total pixels after shrinking are equivalent to a 1300 × 1300 image. | |
| auto | ||||
| xhigh | 1690000 | The image is automatically scaled while maintaining its aspect ratio, and the total pixels after scaling are equivalent to a 1300 × 1300 image | The image is automatically scaled while maintaining its aspect ratio, and the total pixels after scaling are equivalent to a 1300 × 1300 image |
The usage limits of the DeepSeek-V4.1-Flash model for image understanding are as follows:
| Usage limit | Requirement |
|---|---|
| Single image size | 50 MB |
| Total request body size | 64 MB |
| Number of images per request | 3600 |
| Total size of image files passed by URL in a single request | 200 MB |
Video Understanding
For video understanding, the DeepSeek-V4.1-Flash model uses the same frame sampling strategy as doubao-seed-2.0 and later models, as follows:
| Frame sampling strategy | DeepSeek-V4.1-Flash model |
|---|---|
| Tokens per frame | Supports continuous dynamic adjustment of tokens within the [64, 384] range, with a default value of 384.• Maximum tokens per frame, used to control the upper limit of single-frame quality. • Minimum tokens per frame, used to control the lower limit of single-frame compression. The default value is 64. |
| Maximum pixels per frame | Maximum tokens per frame × 42 × 42 [110K, 670K] |
| Number of sampled frames | • Frame sampling range: [16 frames, 1280 frames]. • Maximum number of sampled frames: 80 × 1024 tokens ÷ 64 tokens/frame = 1280 frames |
| Maximum video tokens | Defaults to 81920 (80 × 1024). |
| Frame sampling scheme | Note • It is recommended to evaluate the output quality and adjust the fps field configuration or video duration as needed. • If the video binary data does not contain frame count encoding information, frames are sampled evenly according to fps. If the maximum video token amount in a single request exceeds the model's maximum input length, an error is returned; if not, no error is returned, but the video input tokens may exceed 80k. • When fps is too high or the video is too long: if the number of processed frame images exceeds 1280, 1280 frames are sampled evenly at 64 tokens per frame image with a time interval of video duration/1280.• When fps is too low or the video is too short: if the number of processed frame images is less than 16, the sampling scheme is as follows: - Total video frames ≥ 16: sample 16 frames evenly. - Total video frames < 16: sample all frames of the video. |
How video understanding works:
The core approach to video processing is "structured concatenation of frames and timestamps", with the following rules:
- For each frame image obtained by sampling the video, a timestamp text is inserted before it. The timestamp format processed by the DeepSeek-V4.1-Flash model is
[<timestamp> second]. For example:[4.0 second],[4.5 second]. - The result is an ordered sequence of "timestamp + image", through which the model understands the temporal logic and content changes of the video.
Reasoning Length Adjustment: reasoning_effort/reasoning.effort/output_config.effort
Support for the DeepSeek-V4.1-Flash model in this platform's APIs is as follows:
| API | Chat API | Messages API | Responses API |
|---|---|---|---|
| Reasoning length parameter | reasoning_effort | output_config.effort | reasoning.effort |
| Default value | high | high | high |
| Compatible mapping rules | Supported levels: none, low, high, max• Passing none: thinking is turned off• Passing minimal: mapped to low• Passing medium / xhigh: mapped to high• Passing ultra: mapped to max | Consistent with the Chat API | Supported levels: none, low, high, max• Passing none: thinking is turned off• Passing minimal: mapped to low• Passing medium / xhigh: mapped to high |
Maximum Output Length: max_tokens/max_completion_tokens/max_output_tokens
Support for the maximum output length parameters of the DeepSeek-V4.1-Flash model is as follows:
| API | Chat API | Responses API |
|---|---|---|
| Parameter | max_tokens/max_completion_tokens | max_output_tokens |
| Rule | Controls the length of the model answer and the model reasoning content; no error is returned when the length is exceeded. Note: If the reasoning content exceeds max_tokens, thinking will be truncated. | Controls the length of the model answer and the model reasoning content; no error is returned when the length is exceeded. |
| Default value | 128K | 128K |