> ## Documentation Index
> Fetch the complete documentation index at: https://docs.os3compute.com/llms.txt
> Use this file to discover all available pages before exploring further.

# DeepSeek V4.1 Flash 1017

> Ids, context, features, the fast id, and what differs from DeepSeek's API.

DeepSeek's original V4.1 Flash checkpoint (2026-10-17), never another model in its place. Preview: no uptime promise yet.

| Model id | Context | Supports | Runs in |
| - | - | - | - |
| `deepseek/deepseek-v4.1-flash-1017` | 1,048,512 tokens | Tools, thinking, images | United States only |
| `deepseek/deepseek-v4.1-flash-1017-fast` | 1,048,512 tokens | Tools, thinking, images | United States only |

## Features

| | |
| - | - |
| APIs | `/v1/chat/completions`, `/v1/messages` and `/v1/responses`, streamed or not. |
| Context | Prompt plus answer. A prompt that does not fit is refused, never cut. |
| Output | `max_tokens` up to 393,216. Default 8,192; 65,536 when thinking; 131,072 at `max`. |
| Thinking | `reasoning_effort`: `none`, `low`, `high` (default) or `max`. Returned in `reasoning_content`; send it back on tool turns. |
| Tools | Your own functions. |
| Images | Up to 600 a request. |
| Long prompts | Over 200,000 tokens: allow 120 s to the first token. |

## Normal and fast

The `-fast` id is the same model on a less crowded lane, for when the normal lane is busy.

* The fast lane runs 2 requests at once per GPU server, at most 2 yours.
* When it is full, the normal lane answers if it has room, else you get a 429 or 503.
* The `x-os3-lane` header says which lane answered, `fast` or `bulk` (normal), and so which price you pay.
* A conversation on the normal lane stays there until idle for about 45 seconds.
* Prompts over about 262,144 tokens go to the normal lane, so both ids take the full context.

## Differences from DeepSeek's API

* Context is 1,048,512 tokens, 64 fewer than DeepSeek's window.

- `max_completion_tokens` works as `max_tokens`.
- `json_object` responses are always an object.
- Image `detail: "low"` takes images up to 512 px.
- Image URLs: about 36 MiB and 300 seconds a request, public hosts only.
- Not offered: the Files API, web search, URL reading, FIM.
- `temperature: 0` is not deterministic.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.