Ollama's OpenAI-compatible /v1 silently drops think: false

Developers found that Ollama's OpenAI-compatible /v1 endpoint silently drops the think: false flag for Qwen3-family models. The model then spends its entire num_predict budget on reasoning and returns empty content; the workaround is switching to the native /api/chat endpoint.

• /v1 drops think: false, so Qwen3 returns empty content
• The whole num_predict budget goes to internal reasoning
• Workaround is switching to the native /api/chat endpoint
• The switch requires code refactoring and different rate limits