Ollama's implementation for gpt-oss is poor.