What Is DeepSeek V4-Pro? Models, API, Function Calling & JSON Output Explained • AI / API

What Is DeepSeek V4-Pro? Models, API, Function Calling & JSON Output Explained

You're wiring a DeepSeek Agent and the docs mentiondeepseek-v4-pro,deepseek-v4-flash, and the retireddeepseek-chat. Function Calling sometimes returnsargumentsthat isn't valid JSON—while JSON Mode and Strict Schema follow different paths?

DeepSeek V4-Prois the flagship MoE model in the DeepSeek V4 family, GA since 2026-08-13 (build DeepSeek-V4-Pro-0813). The API surface is unchanged:base_urlstayshttps://api.deepseek.com, setmodeltodeepseek-v4-pro.

It's OpenAI Chat Completions–compatible and also offers an Anthropic-style interface; Function Calling, JSON output, and three thinking effort levels share the same API.

This article covers:

• What V4-Pro and V4-Flash are and how to choose
• A minimal Chat Completions request
• The full Function Calling tool loop
• JSON Mode vs beta Strict Schema
• How JSONNote helps debug model JSON

What is DeepSeek V4-Pro

DeepSeek V4 launched via API on 2026-04-24 as a Mixture-of-Experts stack. V4-Pro is the largest variant: ~1.6T total params, ~49B active per forward pass. The siblingdeepseek-v4-flashis lighter (284B / 13B active) for latency- and cost-sensitive workloads.

The 2026-08-13 GA release significantly boosts Agent skills—official benchmarks like Terminal Bench 2.1 and Toolathlon-Verified report production-grade gains. For developers the practical changes are:

• Multi-step tool calls are steadier, with fewer empty loops before answering
• Native OpenAIResponses APIformat support (Codex-friendly workflows)
• Thinking mode addslow / high / maxthree effort levels

If you already use the OpenAI SDK with another model, switching to V4-Pro usually means changing onlybase_url,api_key, andmodel.

V4 lineup and legacy names

New projects and CI scripts should usedeepseek-v4-proordeepseek-v4-flashdirectly—legacy aliases were retired on 2026-07-24.

API essentials

Minimal Chat Completions request

DeepSeek API matches OpenAI Chat Completions. Python example (install theopenaipackage):

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_DEEPSEEK_API_KEY",
    base_url="https://api.deepseek.com",
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "用 JSON 列出三个 HTTP 状态码及含义"},
    ],
)
print(resp.choices[0].message.content)

Three easy pitfalls

1. 1
   Stateless API

   Every request must include the fullmessageshistory. The server won't remember prior tool results for you.

2. 2
   Prompt Cache

   Responsesusagemay includeprompt_cache_hit_tokens. Repeated calls with the same prefix hit cache and cost less.

3. 3
   Peak / off-peak pricing

   Since 2026-08-16, pricing varies by time slot—off-peak is about half of peak. Batch jobs can be scheduled accordingly.

Function Calling end-to-end

Function Calling lets the model return structured tool invocations—query a database, call HTTP, run math—instead of guessing. V4-Pro uses OpenAI-styletoolsarrays.

1. Define tools (JSON Schema)

{
  "type": "function",
  "function": {
    "name": "get_order",
    "description": "按订单号查询订单状态",
    "parameters": {
      "type": "object",
      "properties": {
        "order_id": { "type": "string", "description": "例如 ORD-10442" }
      },
      "required": ["order_id"]
    }
  }
}

2. Model returns tool_calls

When the model wants a tool,finish_reasonis"tool_calls",message.contentis usually empty—the real instruction lives in thetool_callsarray:

{
  "choices": [{
    "message": {
      "role": "assistant",
      "content": "",
      "tool_calls": [{
        "id": "call_0_f1c29a44",
        "type": "function",
        "function": {
          "name": "get_order",
          "arguments": "{\"order_id\": \"ORD-10442\"}"
        }
      }]
    },
    "finish_reason": "tool_calls"
  }]
}

Note:argumentsis astringyou mustjson.loadsyourself. Official docs warn the model doesn't always emit valid JSON and may hallucinate fields outside your schema.

3. Execute and return a tool message

import json

messages = [{"role": "user", "content": "查一下 ORD-10442 的状态"}]
tools = [/* 上面的 get_order 定义 */]

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=messages,
    tools=tools,
)
msg = resp.choices[0].message

if msg.tool_calls:
    messages.append(msg)  # 保留 assistant 的 tool_calls
    call = msg.tool_calls[0]
    try:
        args = json.loads(call.function.arguments)
    except json.JSONDecodeError:
        args = {}  # 降级或重试
    result = get_order(**args)  # 你的业务函数
    messages.append({
        "role": "tool",
        "tool_call_id": call.id,
        "content": json.dumps(result, ensure_ascii=False),
    })
    # 再次请求，模型基于真实数据生成最终回答
    final = client.chat.completions.create(
        model="deepseek-v4-pro",
        messages=messages,
        tools=tools,
    )

tool_choice behavior

max_tokensToo small aargumentstruncates tool-callfinish_reasonbecomeslengthwith incomplete arguments. Agent workloads need generous completion budgets.

JSON output and Strict mode

JSON Mode (response_format)

When you only need JSON text—not external tools—use JSON Mode:

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "只输出合法 JSON，不要 markdown 代码块"},
        {"role": "user", "content": "生成一个含 name、age 的用户对象"},
    ],
    response_format={"type": "json_object"},
)
raw = resp.choices[0].message.content
data = json.loads(raw)  # 仍建议 try/except

DeepSeek raised internal JSON parse rates from 78% to 85% in the V2 era; regex cleanup can reach 97%. V4 improves further, butalways validate in your application—don't treat a successfuljson.loadsas proof the schema is correct.

Strict Schema (beta)

For Function Calling that must match JSON Schema exactly, enable beta Strict mode:

• base_urlChangehttps://api.deepseek.com/beta
• Set"strict": true
• Everyobjectobject'spropertiesmust list all keys inrequired, withadditionalProperties: false

Strict validates atrequest time—invalid schemas return 400 before generation. Good for catching schema bugs at deploy time.

JSON Mode vs Function Calling

Three thinking effort levels

V4-Pro and V4-Flash thinking modes supportlow / high / maxeffort tiers (seeofficial API docsThinking Mode section):

• low: simple Q&A, formatting, short summaries
• high: daily Agents, multi-step reasoning
• max: complex code generation, long planning

Thinking costs more tokens and latency. For “turn this text into JSON,” non-thinking mode + JSON Mode is usually cheaper.

V4-Pro vs V4-Flash

Both share the same API—A/B tests only need amodelchange. A common rollout: prove the tool chain on Flash, upgrade to Pro when needed.

Debug model JSON with JSONNote

Integrating V4-Pro means lots of time asking “does this JSON parse, does the schema match?” JSONNote runs locally in the browser—your API key and response bodies never upload.

1. 1
   Format & validate

   Pastemessage.contentortool_calls[].function.argumentsintoJSON FormatterorJSON Validatorto see syntax errors instantly.

2. 2
   Check against Schema

   In theJSON Schematool, paste your function parameters schema and model output to verify fields.

3. 3
   Diff two calls

   After tweaking prompts or switching Pro / Flash, useJSON Diffto compare structured outputs.

4. 4
   Share with teammates

   UseURL Hash sharingto encode failing JSON and tool state in a link—recipients reproduce locally; data never hits our servers.

FAQ

What is the DeepSeek V4-Pro API model name?

deepseek-v4-pro.base_urlishttps://api.deepseek.com, auth matches OpenAI (Authorization: Bearer …).

How do I choose between V4-Pro and V4-Flash?

Complex Agents, long context, high tool accuracy → Pro. High concurrency, cost-sensitive, standard tasks → Flash. Both support Function Calling and JSON output.

What if Function Calling returns malformed JSON?

Wrap parsing intry/except json.JSONDecodeError; validate required fields; retry or fall back to JSON Mode; use beta Strict when you need strict adherence.

Do I still need my own schema with JSON Mode?

JSON Mode only guarantees a JSON object—not that fields match your business schema. Enforce shape via system prompt, post-validation, or Function Calling + Strict.

Can I still use deepseek-chat?

No. Legacy names retired 2026-07-24. Migrate todeepseek-v4-proordeepseek-v4-flash.

How does Function Calling differ from OpenAI gpt-4o?

Request/response fields and tool loops are largely compatible. Differences: pricing, context length, thinking tiers, and DeepSeek's Strict beta path. Regression-testargumentsparsing andmax_tokensbudgets when migrating.

Summary

DeepSeek V4-Pro is essentially:

OpenAI-compatible API + flagship MoE + native Function Calling / JSON output + optional thinking effort.

Setmodeltodeepseek-v4-pro, write tool loops the OpenAI way, but remember:argumentsis model-generated—parse and validate on your side. JSONNote's formatter, Schema, and Diff tools help debug structured output locally.

← Back to blog