litellm - 💡(How to fix) Fix [Bug]: DeepSeek V4 Pro (deepseek-v4-pro) fails in multi-turn conversations - reasoning_content stripped from assistant messages [1 participants]

What Happened?

When using deepseek/deepseek-v4-pro (or the Anthropic-compatible endpoint anthropic/deepseek-v4-pro with api_base: https://api.deepseek.com/anthropic) through LiteLLM proxy, the first turn succeeds but every subsequent turn fails with:

litellm.BadRequestError: DeepseekException - {"error":{"message":"The `reasoning_content` in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}

Expected behavior: Multi-turn conversations should work correctly with DeepSeek V4 Pro.

Root Cause Analysis

DeepSeek V4 Pro enables thinking mode by default and has a new API requirement compared to older DeepSeek reasoning models:

• Old behavior (deepseek-reasoner / R1): reasoning_content must NOT be passed back in subsequent turns — passing it causes a 400 error.
• New behavior (deepseek-v4-pro): reasoning_content MUST be passed back in subsequent turns — omitting it causes a 400 error.

LiteLLM currently handles the old R1 behavior correctly: in litellm/types/utils.py (around line 1224), the Message.__init__ method deletes reasoning_content when it is None during serialization. This means reasoning_content from the first-turn response is stripped before being sent as the assistant message in the second turn.

This worked for R1 but breaks V4 Pro, which requires the field to be preserved.

Steps to Reproduce

config.yaml:

model_list:
  - model_name: deepseek-v4-pro
    litellm_params:
      model: deepseek/deepseek-v4-pro
      api_key: os.environ/DEEPSEEK_API_KEY

curl — first turn (succeeds):

curl http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [{"role": "user", "content": "What is 1+1?"}]
  }'

curl — second turn (fails with 400):

curl http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $LITELLM_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [
      {"role": "user", "content": "What is 1+1?"},
      {"role": "assistant", "content": "2"},
      {"role": "user", "content": "And 2+2?"}
    ]
  }'

The second turn fails because the assistant message is missing reasoning_content — LiteLLM strips it when serializing the response from turn 1.

Relevant Log Output

litellm.BadRequestError: DeepseekException - {"error":{"message":"The `reasoning_content` in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}. Received Model Group=deepseek-v4-pro

Suggested Fix

For deepseek-v4-pro (and potentially any DeepSeek model with thinking mode enabled), LiteLLM should:

1. Preserve reasoning_content in the assistant message when building the conversation history for subsequent requests, OR
2. Detect that the model requires thinking mode and automatically inject reasoning_content back into the messages array before forwarding to DeepSeek.

Component

• Proxy
• SDK (litellm Python package)

Version

• LiteLLM: v1.82.3 (also reproduced conceptually against v1.83.12-nightly based on source code review)
• DeepSeek V4 Pro API (released 2026-04-24)