DeepSeek V3 → V4 Migration Guide

• 2026-04-24: V4 preview released — V4-Pro (1.6T/49B) and V4-Flash (284B/13B) launched together, MIT open source.
• Now: Migration window — Legacy names still work, but the official guidance is to switch early to avoid last-minute deployment pressure.
• 2026-07-24 15:59 UTC: Hard deadline — deepseek-chat / deepseek-reasoner deprecated.

Flash vs Pro core difference: Flash is lightweight (284B/13B) with 2500 concurrency; Pro is flagship (1.6T/49B) with 500 concurrency but stronger coding and reasoning. Use Flash for daily Q&A; Pro for agent coding and long-chain reasoning.

In other words, migration work = change model parameter + adapt thinking mode parameters (if coming from deepseek-reasoner). Everything else stays.

Python + OpenAI SDK (most common)

Before

# 迁移前
from openai import OpenAI

client = OpenAI(
    api_key="<DeepSeek API Key>",
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-chat",  # ❌ 旧名
    messages=[{"role": "user", "content": "Hello"}],
)

After

# 迁移后
from openai import OpenAI

client = OpenAI(
    api_key="<DeepSeek API Key>",
    base_url="https://api.deepseek.com"  # 不变
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",  # ✅ 新名
    messages=[{"role": "user", "content": "Hello"}],
)

Anthropic SDK (Claude Code / Cursor / etc.)

Before

# 迁移前（如果你之前用第三方中转接 DeepSeek）
# Claude Code settings.json
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_API_KEY": "<你的 Key>"
  }
}

After

# 迁移后：Claude Code 直接用 V4
{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_API_KEY": "<你的 Key>",
    "ANTHROPIC_MODEL": "deepseek-v4-flash"  // 或 deepseek-v4-pro
  }
}

OpenAI SDK: Enable thinking mode

from openai import OpenAI

client = OpenAI(
    api_key="<DeepSeek API Key>",
    base_url="https://api.deepseek.com"
)

response = client.chat.completions.create(
    model="deepseek-v4-flash",
    messages=[{"role": "user", "content": "9.11 and 9.8, which is greater?"}],
    reasoning_effort="high",  # high / max
    extra_body={"thinking": {"type": "enabled"}},
)

# 思考过程在 reasoning_content，最终答案在 content
reasoning = response.choices[0].message.reasoning_content
answer = response.choices[0].message.content

Anthropic SDK: Thinking intensity

# Anthropic 兼容端点
{
  "model": "deepseek-v4-flash",
  "thinking": {"type": "enabled"},
  "output_config": {"effort": "high"}
}

Multi-turn: reasoning_content handling

V4 has a multi-turn detail: without tool calls, previous reasoning_content doesn't need to be passed back (API ignores it); with tool calls, it must be fully passed back, otherwise a 400 error.

# 推荐写法：直接 append 整个 message 对象，reasoning_content 自动包含
messages.append(response.choices[0].message)
# 而不是手动复制 content / reasoning_content / tool_calls 字段

1M context application tests → long-context page.

Source: DeepSeek official API docs, 2026-07-05. Cache hit applies to repeated prefix requests (multi-turn, agent loops). Flash concurrency limit 2500; Pro 500.

• Grep code for deepseek-chat — Replace with deepseek-v4-flash (non-thinking) or deepseek-v4-pro (when more capability is needed)
• Grep code for deepseek-reasoner — Replace with deepseek-v4-flash, add extra_body={'thinking': {'type': 'enabled'}} + reasoning_effort='high'
• Verify temperature params are ignored — In thinking mode, temperature/top_p/presence_penalty/frequency_penalty have no effect. If your logic depends on these, switch to Pro or rewrite logic.
• Test 1M context (if needed) — Old code may assume 128K ceiling. If you have chunking logic, you can simplify or remove it.
• Test multi-turn reasoning_content — Tool call flows must include reasoning_content, otherwise 400 error.
• Update monitoring and alerts — Search your monitoring dashboards for deepseek-chat / deepseek-reasoner references, change to v4 model names.
• Rollback plan — Before production rollout, ensure V3 API key still has balance (V4 doesn't need a new key, uses old one). If V4 has issues, temporarily rolling back to deepseek-chat still works until 7-24.