Sounds like a problem with your provider, or with how your harness sends requests.

All DeepSeek models have 384k maximum output tokens:

https://api-docs.deepseek.com/quick_start/pricing