DeepSeek-V4-Pro-0813 Now Live on SiliconFlow

TL;DR

• What launched: DeepSeek-V4-Pro-0813, the official release of DeepSeek-V4-Pro, is now live on SiliconFlow. It supersedes the preview version with substantially enhanced agent capabilities and improvements that are especially pronounced in production environments.

• Key features: A 1M-token context window, reasoning, tool calling, prefix completion, a Mixture-of-Experts architecture, and a DSpark speculative decoding module.

• Cost on SiliconFlow: $1.32/M input tokens · $3.96/M output tokens · $0.44/M cached input tokens.

• Get started: Try it in the SiliconFlow Model Library or connect through the OpenAI-compatible Chat Completions API.

DeepSeek-V4-Pro-0813 is now available on SiliconFlow for developers building coding agents, tool-using assistants, and production workflows that require sustained execution across multiple steps.

The release builds on the DeepSeek-V4-Pro Preview model structure and adds a DSpark speculative decoding module. More importantly, DeepSeek’s August 13 update reports substantial improvements across terminal operation, repository understanding, software engineering, tool use, and workflow automation benchmarks.

DeepSeek positions the official release around significantly enhanced agent capabilities, with performance improvements described as especially pronounced in production environments.

From Preview to the Official Release

The DeepSeek-V4 Preview introduced a model family designed around a million-token context, advanced reasoning, coding, and agentic tasks.

On July 31, DeepSeek released an updated V4-Flash and reported substantial gains over V4-Pro Preview across multiple agent benchmarks, and DeepSeek stated that the official release of V4-Pro would follow.

On August 13, DeepSeek announced the official release of V4-Pro across its app, web, and API services. The update highlighted significantly enhanced agent capabilities, with improvements described as especially pronounced in production environments.

That official release is now available on SiliconFlow as DeepSeek-V4-Pro-0813.

DeepSeek-V4-Pro-0813 on SiliconFlow

Why DeepSeek-V4-Pro-0813 Stands Out

Stronger Agent Performance in Production Environments

Its reported results cover:

• Terminal-based task execution

• Repository-level code understanding

• End-to-end software engineering

• Tool selection and invocation

• Full-stack development

• Workflow automation

Reasoning and Tool Use in the Same Workflow

DeepSeek-V4-Pro-0813 supports both reasoning and tool calling. DeepSeek’s official thinking-mode guide demonstrates how the model can use the result of one tool call to prepare the next action in a multi-step workflow.

Million-Token Context with DSpark Speculative Decoding

The 1M-token context window allows DeepSeek-V4-Pro-0813 to work with substantially more information within a single task.

Depending on the application, that context can include:

• Large code repositories

• Long technical documents

• Multi-stage tool results

• Extended conversation histories

• Specifications, logs, and supporting evidence

• Long-running Agent task state

The release also attaches a DSpark speculative decoding module to the DeepSeek-V4-Pro Preview model structure.

Benchmark Performance

According to DeepSeek’s official update log, the August 13 release improves on V4-Flash-0731 across a range of agent benchmarks.

DeepSeek-V4-Pro-0813 scores higher than DeepSeek-V4-Flash-0731 on all eight benchmarks listed above.

Example: A Multi-Step Tool-Calling Agent

DeepSeek’s official thinking-mode guide includes a tool-calling example using deepseek-v4-pro.

The Task

The user asks:

How’s the weather in Hangzhou tomorrow?

Answering correctly requires more than generating a weather statement. The Agent must determine the current date, calculate which date “tomorrow” refers to, and then request the weather for the correct city and date.

Why the Task Requires Multiple Steps

The example gives the model access to two tools:

1. A tool that returns the current date

2. A tool that returns the weather for a specified city and date

The weather tool cannot be called correctly until the date tool has returned its result.

Execution Flow

The model completes the task in the following sequence:

1. Identifies that it first needs the current date

2. Calls the date tool

3. Calculates the date corresponding to “tomorrow”

4. Calls the weather tool with Hangzhou and the calculated date

5. Reads the returned weather data

6. Produces a final user-facing response

Additional Prompting

The user message in the official example is a single sentence: “How’s the weather in Hangzhou tomorrow?” The model receives the user’s request together with the function schemas for the available tools.

DeepSeek’s official thinking-mode guide explains how the relevant assistant messages should be carried forward across follow-up requests involving tool calls. 

Final Result

In DeepSeek’s official example, the mocked weather function returns “Cloudy 7–13°C” after the model resolves the date dependency and completes the two tool calls.

What the Example Demonstrates

The example demonstrates a complete dependency-aware tool workflow:

Identify the dependency → call the first tool → use its result → call the next tool → deliver the final answer

The official example shows the model using the date tool’s result to construct the arguments for the subsequent weather tool call.

Run DeepSeek-V4-Pro-0813 with Your Existing Tools

A simple way to evaluate DeepSeek-V4-Pro-0813 is to try it on a task from your own workflow.

Try It First, No Setup

Open the SiliconFlow Model Library, select DeepSeek-V4-Pro-0813, and launch it in the Playground.

Useful first tests include:

• Reviewing a multi-file implementation plan

• Identifying dependencies across a code repository

• Designing a tool-calling workflow

• Analyzing a long technical document

• Comparing two implementation approaches

• Breaking a production issue into executable investigation steps

Plug It into Your Tools

SiliconFlow APIs use an OpenAI-compatible interface, allowing DeepSeek-V4-Pro-0813 to connect to development tools and Agent frameworks that support custom providers.

You only need three pieces of information:

Developers can use the model with coding assistants, terminal agents, custom Agent systems, and existing applications built around OpenAI-compatible Chat Completions.

Get Started Immediately

Explore

Open DeepSeek-V4-Pro-0813 in the SiliconFlow Model Library and test the model in the Playground.

Integrate

Use the SiliconFlow Chat Completions API with your SiliconFlow API Key and the new Model String.

import os
import requests

api_key = os.environ["SILICONFLOW_API_KEY"]

response = requests.post(
    "https://api.siliconflow.com/v1/chat/completions",
    headers={
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json",
    },
    json={
        "model": "deepseek-ai/DeepSeek-V4-Pro-0813",
        "messages": [
            {
                "role": "system",
                "content": (
                    "You are a careful software engineering assistant. "
                    "Analyze dependencies before proposing changes."
                ),
            },
            {
                "role": "user",
                "content": (
                    "Review this implementation task, propose an execution plan, "
                    "and identify the main technical risks."
                ),
            },
        ],
        "stream": False,
        "max_tokens": 4096,
    },
    timeout=120,
)

response.raise_for_status()

result = response.json()
print(result["choices"][0]["message"]["content"])

import os
import requests

api_key = os.environ["SILICONFLOW_API_KEY"]

response = requests.post(
    "https://api.siliconflow.com/v1/chat/completions",
    headers={
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json",
    },
    json={
        "model": "deepseek-ai/DeepSeek-V4-Pro-0813",
        "messages": [
            {
                "role": "system",
                "content": (
                    "You are a careful software engineering assistant. "
                    "Analyze dependencies before proposing changes."
                ),
            },
            {
                "role": "user",
                "content": (
                    "Review this implementation task, propose an execution plan, "
                    "and identify the main technical risks."
                ),
            },
        ],
        "stream": False,
        "max_tokens": 4096,
    },
    timeout=120,
)

response.raise_for_status()

result = response.json()
print(result["choices"][0]["message"]["content"])

import os
import requests

api_key = os.environ["SILICONFLOW_API_KEY"]

response = requests.post(
    "https://api.siliconflow.com/v1/chat/completions",
    headers={
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json",
    },
    json={
        "model": "deepseek-ai/DeepSeek-V4-Pro-0813",
        "messages": [
            {
                "role": "system",
                "content": (
                    "You are a careful software engineering assistant. "
                    "Analyze dependencies before proposing changes."
                ),
            },
            {
                "role": "user",
                "content": (
                    "Review this implementation task, propose an execution plan, "
                    "and identify the main technical risks."
                ),
            },
        ],
        "stream": False,
        "max_tokens": 4096,
    },
    timeout=120,
)

response.raise_for_status()

result = response.json()
print(result["choices"][0]["message"]["content"])

Example Configuration

• Use the exact Model String: deepseek-ai/DeepSeek-V4-Pro-0813.

• Set stream to false when the application should receive the complete response in a single JSON object.

• Set stream to true when the application should process tokens incrementally through Server-Sent Events.

• Set max_tokens according to the expected response length. The example uses 4096; ensure that the input and generated tokens remain within the model’s context window.

• Store the SiliconFlow API Key in an environment variable rather than placing it directly in application code.

• Additional parameters should be added only after confirming their current model-specific support in the SiliconFlow API documentation or the SiliconFlow Model Library.

• Business or Sales Inquiries →

• Join our Discord community now →

• Follow us on X for the latest updates →

• Explore all available models on SiliconFlow →