DeepSeek V4 Flash 0731 Now Live on SiliconFlow

TL;DR

• DeepSeek-V4-Flash-0731 is now live on SiliconFlow: The official release of DeepSeek V4 Flash — same size and architecture as the preview, re-post-trained for a significant jump in agentic capability.

• Why it matters: Independent testing backs up DeepSeek's own claim. On Artificial Analysis's Intelligence Index, 0731 scores 50 — six points ahead of DeepSeek's flagship model — and lands just one point behind GLM-5.2. DeepSeek's own published agentic benchmark suite shows the same story in more depth: 0731 beats DeepSeek-V4-Pro on every listed benchmark, while closing the gap with Opus 4.8.

• What it costs: $0.028 / M tokens for cache reads, $0.13 / M input tokens, and $0.28 / M output tokens on SiliconFlow.

• Get started: SiliconFlow provides an OpenAI-compatible API for 0731 — model string deepseek-ai/DeepSeek-V4-Flash-0731 — so it drops straight into Claude Code, Cline, OpenCode, Hermes Agent, Codex via CC Switch, and your existing tools in seconds.

DeepSeek released V4 Flash as a preview version in April 2026. Three months later, 0731 is the official release: same 284B-total, 13B-activated architecture, re-post-trained specifically for agentic work. The jump is large enough that DeepSeek's own benchmark suite shows 0731 beating DeepSeek V4 Pro on every listed benchmark and significantly narrowing the gap with Opus 4.8.

DeepSeek V4 Flash 0731 pairs Flash-tier speed and pricing with agentic capability that closes in on the flagship — and it's now live on SiliconFlow. Here's what you are working with:

Why DeepSeek V4 Flash 0731 Stands Out

Enhanced Agentic Capabilities

DeepSeek did not make V4 Flash larger for the official release. V4 Flash 0731 retains the same model size and architecture as the preview; the upgrade comes entirely from a new round of post-training focused on agentic work—from planning and tool use to multi-step execution and recovery from failed attempts.

The result is a broad leap across DeepSeek's agentic benchmark suite. V4 Flash 0731 outperforms both DeepSeek V4 Flash and DeepSeek V4 Pro on every listed test, with gains spanning repository-level coding, terminal work, tool use, cybersecurity, automation, and full-stack development. It still trails Opus 4.8 overall, but the gap is now substantially narrower—all without moving the model out of Flash-class size or pricing.

Source: DeepSeek's official Change Log and the DeepSeek-V4-Flash-0731 model card. Opus 4.8 leads on every row above — this table is about how far 0731 closes that gap against DeepSeek's own flagship preview and its predecessor, not about topping the field.

Flash Speed and Pricing

Despite the leap in agent capabilities, V4 Flash 0731 retains the speed and pricing of a Flash-tier model. On SiliconFlow, it remains $0.028 / M tokens for cache reads, $0.14 / M input tokens, and $0.28 / M output tokens.

Artificial Analysis's independent tracking reinforces that positioning. Among the 36 SiliconFlow deployments it tracks, V4 Flash 0731 Max ranks second in intelligence with an Intelligence Index score of 52, delivers the fastest output at 136 tokens per second, and is the model with the lowest blended price at $0.1 per million tokens.

Three Reasoning Levels, One Model

Not every task needs the same amount of reasoning.

DeepSeek V4 Flash 0731 now supports three reasoning-effort levels:

Use low while exploring and iterating; move to high or max when the task needs deeper planning or a longer autonomous run — without switching models.

Run DeepSeek V4 Flash 0731 with Your Familiar Tools

Benchmarks and backstory can only tell you so much — the fastest way to know if 0731 changes your workflow is to run it yourself. That only takes minutes.

Try it first, no setup

Open the SiliconFlow Playground and chat with DeepSeek-V4-Flash-0731 directly in your browser. Try a coding question, paste in a self-contained bug, or ask the model to work through a technical problem step by step. To see the difference directly, run the same prompt against DeepSeek-V4-Pro or GLM-5.2 side by side.

Plug into your tools

SiliconFlow's API is OpenAI-compatible, so 0731 drops straight into any tool that supports a custom provider — same SDK, same code, just the SiliconFlow base URL and model string.

What you need to connect

💡 One tip for Codex: Keep CC Switch's local routing enabled while using the SiliconFlow provider. The routing layer needs to stay active to translate requests and streaming responses between the two protocols.

Get Started Immediately

1. Try it: Start using DeepSeek-V4-Flash-0731 on SiliconFlow — run queries immediately, pay only for usage.

2. Integrate: Use our OpenAI-compatible API. Explore the full API specifications in the SiliconFlow API documentation. DeepSeek recommends temperature = 1.0 and top_p = 1.0 for general use.

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "stream": False,
    "messages": [
        {
            "role": "user",
            "content": "A test in this repo is failing intermittently. Investigate the failure, form a hypothesis, and propose a fix — explain your reasoning before changing code."
        }
    ],
    "max_tokens": 4096,
    "temperature": 1.0,
    "top_p": 1.0
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.request("POST", url, json=payload, headers=headers)

print(response.text)

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "stream": False,
    "messages": [
        {
            "role": "user",
            "content": "A test in this repo is failing intermittently. Investigate the failure, form a hypothesis, and propose a fix — explain your reasoning before changing code."
        }
    ],
    "max_tokens": 4096,
    "temperature": 1.0,
    "top_p": 1.0
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.request("POST", url, json=payload, headers=headers)

print(response.text)

import requests

url = "https://api.siliconflow.com/v1/chat/completions"

payload = {
    "model": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "stream": False,
    "messages": [
        {
            "role": "user",
            "content": "A test in this repo is failing intermittently. Investigate the failure, form a hypothesis, and propose a fix — explain your reasoning before changing code."
        }
    ],
    "max_tokens": 4096,
    "temperature": 1.0,
    "top_p": 1.0
}
headers = {
    "Authorization": "Bearer <token>",
    "Content-Type": "application/json"
}

response = requests.request("POST", url, json=payload, headers=headers)

print(response.text)

• Business or Sales Inquiries →

• Join our Discord community now →

• Follow us on X for the latest updates →

• Explore all available models on SiliconFlow →