DeepSeek V4.1 Flash

DeepSeek V4.1 Flash 1017 is the October 17, 2026 update behind deepseek-flash, with unchanged token prices and MIT-licensed weights.

What changed in V4.1 Flash 1017
Release changes, benchmark conditions, official demonstrations and open-weight versus hosted access.
Official project demonstrations
DeepSeek's recorded Forbidden City, racing game, Godot Portal2 and Valley of Ages examples. Selected coding projects, not native video generation or independent model tests.
Search and Fetch billing
Separate request-based charges; the temporary server-side integration exemption ends November 1, 2026 at 00:00 Beijing time.
AI Model API
Compare provider IDs, inputs, outputs, prices, and access boundaries before selecting an integration.

Language ModelMultimodalCoding AgentOpen WeightsLong Context

DeepSeek API or large-scale self-hosted inference

Off-peak: $0.003/1M cache-hit input, $0.15/1M cache-miss input, $0.60/1M output; peak: $0.006, $0.30 and $1.20 respectively

Commercial

🚀Function Overview

Use the hosted model for coding, reasoning, image understanding and agent workflows, or evaluate the dated 1017 open checkpoint on suitable inference hardware. DeepSeek reports stronger results and shorter tasks than September V4.1; these are vendor claims, not an independent test.

Key Features

Current direct API ID deepseek-flash serves DeepSeek-V4.1-Flash-1017; the ID does not pin a checkpoint
Legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp IDs temporarily route to this model
1M context and up to 384K output; the output ceiling is not the default request budget
Text and image input with text output
Thinking and non-thinking modes with low, high and max effort behavior
JSON output, tool calls, Responses API and Anthropic-compatible Messages
552B-parameter backbone with about 8B active during input prefill and 16B during output decode
Current MIT-licensed checkpoint deepseek-ai/DeepSeek-V4.1-Flash-1017; September weights retain their separate identity
Off-peak pricing of $0.003 cache-hit input, $0.15 cache-miss input and $0.60 output per 1M tokens
Peak pricing of $0.006 cache-hit input, $0.30 cache-miss input and $1.20 output per 1M tokens
Account concurrency limit of 2,500 requests
DeepSeek recommends Harness v0.2.1-rc.1 or later for the updated model
Search and Fetch are separate APIs at $0.75 per 1,000 successful calls each; their charges are additional to model tokens

Use Cases

•Cost-sensitive long-context coding and agent sessions
•Screenshot, chart and document understanding inside tool-using workflows
•Responses- or Anthropic-compatible agent integrations
•Structured JSON and function-calling applications
•Self-hosted research and inference on suitable multi-GPU or distributed hardware
•Migration from the retired V4 Flash and Vision Exp endpoints

⚙️Input Parameters

model

string

For POST https://api.deepseek.com/chat/completions, use deepseek-flash; DeepSeek identifies the current backend as V4.1-Flash-1017.

messages

array

Chat Completions messages with role and content. A text string supports text-only requests; actual image requests require the documented image content blocks.

thinking

object

Optional type enabled or disabled; thinking is enabled by default.

reasoning_effort

string

none disables thinking; low, high and max enable it. Default high. Use the chosen API protocol's documented field shape.

max_tokens

integer

Optional output budget from 1 to 393,216 (384K), within the total context budget. Defaults: 8K non-thinking, 64K thinking, 128K at max effort.

tools

array

Optional function-tool definitions for Chat Completions; the caller executes supplied functions. Standalone Search/Fetch have separate documented endpoints.

💡Usage Examples

Example 1

Input Parameters

{
  "model": "deepseek-flash",
  "messages": [
    {
      "role": "user",
      "content": "Write a Python function that returns the unique items in a list while preserving their original order. Include two edge-case tests."
    }
  ],
  "thinking": {
    "type": "enabled"
  },
  "reasoning_effort": "high",
  "max_tokens": 4096
}

Output Results

Illustrative request for POST https://api.deepseek.com/chat/completions; not executed by GetLLMs. Expect a Chat Completions response containing the generated answer in choices, not a measured quality or latency result.

Quick Actions

Technical Specifications

Hardware Type: DeepSeek API or large-scale self-hosted inference
Commercial Use: Supported
Pricing: Off-peak: $0.003/1M cache-hit input, $0.15/1M cache-miss input, $0.60/1M output; peak: $0.006, $0.30 and $1.20 respectively