DeepSeek V4 Flash

DeepSeek V4 Flash is a low-cost coding-agent model upgraded to DeepSeek-V4-Flash-0731 on July 31, 2026, with a public-beta API, native Responses API and Codex support, plus an official MIT-licensed checkpoint for self-hosting.

Platform: Replicate

Language ModelCoding AgentResponses APILong Context

DeepSeek API or self-hosted multi-GPU

Off-peak: $0.007/1M cache-hit input, $0.22/1M cache-miss input, $0.66/1M output; peak: $0.014, $0.44, and $1.32 respectively

Commercial

🚀Function Overview

A public-beta API and open-weight model for cost-sensitive coding agents, long-context text workflows, structured output, function calling, Responses-compatible clients such as Codex, and suitably provisioned self-hosted inference.

Key Features

Stable request model ID deepseek-v4-flash with backend version DeepSeek-V4-Flash-0731
1M token context length and 384K maximum output
Native stateless Responses API support with function tools and server-side web search
Official Codex integration path for CLI, desktop, and VS Code clients
Thinking enabled by default with low, high, and max effort behavior
OpenAI Chat Completions and Anthropic-compatible API access
Official DeepSeek-V4-Flash-0731 checkpoint released under the MIT License with vLLM and SGLang serving guidance
Off-peak pricing of $0.007 cache-hit input, $0.22 cache-miss input, and $0.66 output per 1M tokens
Peak pricing of $0.014 cache-hit input, $0.44 cache-miss input, and $1.32 output per 1M tokens during the documented weekday UTC windows
Account concurrency limit of 2,500 requests

Use Cases

•Cost-sensitive coding-agent sessions through the DeepSeek API
•Codex workflows using a Responses-compatible custom provider
•Long-context text analysis with tool calls or server-side web search
•Structured JSON or JSON Schema output generation
•Self-hosted inference on suitable multi-GPU hardware when privacy or sustained volume justifies the operational cost
•Controlled comparison with the generally available DeepSeek V4 Pro endpoint
•Text-only workflows that do not require the separate experimental Vision-Exp API model

⚙️Input Parameters

input

string | array

Text or Responses API input items. Supported item types include messages, reasoning, function calls, function outputs, and web-search calls; image and file inputs are not supported.

instructions

string

Optional system-level instruction inserted at the beginning of the Responses API context.

reasoning

object

Optional thinking configuration. DeepSeek maps effort values to non-thinking, low, high, or max behavior.

tools

array

Optional function definitions and the supported server-side web-search tool for Responses API workflows.

text.format

object

Optional text, JSON object, or JSON Schema output configuration.

💡Usage Examples

Example 1

Input Parameters

{
  "model": "deepseek-v4-flash",
  "input": "Review this change plan and return the three highest-risk assumptions.",
  "reasoning": {
    "effort": "high"
  },
  "max_output_tokens": 1200
}

Output Results

A DeepSeek Responses API object. The endpoint is stateless, so resend the required conversation history on later turns and verify the current schema before production use.

Quick Actions

Use NowView Documentation

Technical Specifications

Hardware Type: DeepSeek API or self-hosted multi-GPU
Commercial Use: Supported
Pricing: Off-peak: $0.007/1M cache-hit input, $0.22/1M cache-miss input, $0.66/1M output; peak: $0.014, $0.44, and $1.32 respectively
Platform: Replicate