DeepSeek V4‑Flash Public Beta: Agent Benchmarks Surpass V4‑Pro Preview with Native Responses API Support

Core Upgrades

Public‑beta DeepSeek V4‑Flash keeps the 284B MoE architecture and 1 M context window; all performance gains come from post‑training alignment optimization.

Full‑stack development benchmark DSBench‑FullStack and complex toolset benchmark Toolathlon scores are significantly higher than the V4‑Pro preview.

Multi‑turn long‑chain tool calls show markedly improved stability, with rare call breaks or parameter loss in MCP service scenarios.

Inference latency remains low, preserving Flash’s speed advantage and reducing cost for high‑frequency development scripts and MCP debugging.

Positioning:

V4‑Flash – high‑speed, cost‑effective model for Codex development, MCP debugging, and everyday agent automation.

V4‑Pro (preview released early August) – flagship precision for large‑scale refactoring and complex multi‑step engineering.

Native Responses API Support

Previous DeepSeek releases offered only a Chat Completions endpoint, requiring a proxy to translate to Codex’s Responses protocol, which often caused 404 errors and tool‑parsing failures. V4‑Flash implements the full Responses API specification and includes dedicated Codex adaptations, enabling a single config.toml change to activate the model across all Codex clients without an intermediary service.

Zero‑Proxy Codex Integration Guide

Prerequisites

Install any Codex client (CLI, VS Code plugin, or desktop app).

Create a DeepSeek OpenAI‑style API key (prefix sk) on the DeepSeek Open Platform.

Use the appropriate configuration file path:

Unix: ~/.codex/config.toml Windows:

C:\Users\<em>username</em>\.codex\config.toml

Option 1: One‑Click Auto‑Configuration Script

Run in a macOS/Linux terminal:

bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup.sh)

Or in an elevated Windows PowerShell:

irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex

Execution flow:

Back up existing config.toml and models.json to a safe directory.

Generate a standard models.json containing metadata for deepseek-v4-flash and deepseek-v4-pro.

Write DeepSeek vendor settings, automatically filling the official API base URL and the Responses protocol flag.

Prompt the user to input their DeepSeek API key; syntax is validated before writing.

Provide an interactive menu to switch models or restore the original Codex configuration.

Option 2: Manual Configuration (No Network Script)

Create ~/.codex/models.json with the following content:

{
  "models": [
    {
      "slug": "deepseek-v4-flash",
      "display_name": "DeepSeek-V4-Flash",
      "description": "Public‑beta code‑agent model, native Responses API support for all Codex clients",
      "context_window": 1048576,
      "supports_parallel_tool_calls": true,
      "tool_mode": null,
      "wire_api": "responses",
      "default_reasoning_level": "high",
      "supported_reasoning_level": ["low", "high", "max"],
      "visibility": "list",
      "minimal_client_version": "0.144.0"
    },
    {
      "slug": "deepseek-v4-pro",
      "display_name": "DeepSeek-V4-Pro",
      "description": "Flagship code model (Responses support added in early August)",
      "context_window": 1048576,
      "supports_parallel_tool_calls": true,
      "wire_api": "responses",
      "default_reasoning_level": "high"
    }
  ]
}

Then edit config.toml to include:

# Default model
model = "deepseek-v4-flash"
model_provider = "deepseek"
# Use API key authentication
preferred_auth_method = "apikey"
forced_login_method = "api"
# High reasoning effort balances speed and capability
model_reasoning_effort = "high"
# Path to the model catalog
model_catalog_json = "~/.codex/models.json"

[model_providers.deepseek]
name = "DeepSeek Large Model"
base_url = "https://api.deepseek.com/"
wire_api = "responses"
experimental_bearer_token = "sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxx"

Configuration Validation and Multi‑Terminal Use

Run codex in any project directory; the log should show model: deepseek-v4-flash, confirming successful integration. Example command:

# MCP service call (compatible with Webman MCP Server / Spring MCP)
codex read order MCP tool, write a PHP customer‑query API

In VS Code, the Codex plugin reads the .codex folder automatically; the model appears as a custom entry labeled “DeepSeek‑V4‑Flash”. The desktop client also lists the custom model in its dropdown, enabling command execution, batch file editing, and full tool‑call capabilities.

Reference URL: https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/codex/