gpt-oss-20b HTML tags spoil chat UX

This feels less like a bad model and more like a formatting pipeline mismatch. OpenAI’s own guidance says gpt-oss should be used with the harmony response format, so if HTML is slipping through, the fix is probably in the runtime, template, or output sanitation layer.

• –OpenAI positions gpt-oss-20b as an open-weight reasoning model meant for self-hosted or provider-run inference, not a chat app that can ignore formatting details
• –If you need polished UX, structured outputs plus a markdown renderer is safer than relying on the model to never emit raw HTML
• –Tables are the tricky part: preserve them in markdown or HTML internally, then normalize before display instead of asking the model to “just output markdown”
• –This is a classic open-model tradeoff: you get control and low-latency local deployment, but the integrator owns the last mile of presentation quality
• –The thread is useful because it spotlights a real production concern that benchmark scores won’t show: output hygiene matters as much as raw capability