Building a DeepSeek-R1 Chatbot

I built a chatbot with DeepSeek-R1-Distill-Qwen-1.5B to experiment with a downloadable reasoning model. The interface separates its intermediate output from the final response.

Try the Hugging Face demo or read the source.

Showing the intermediate output

While testing responses, I noticed the model produced reasoning text before its answer. I decided to display it rather than discard it.

Asked about climate change in Malaysia, it produced a structured response but included the Andaman and Nicobar Islands and a “Ch bamboo” initiative I had never heard of. The output sounded convincing despite those mistakes.

Showing the intermediate text helped me inspect the response and compare prompts. It still needed fact-checking.

Hosting the experiment

I used Gradio on a basic Hugging Face Spaces instance with 2 vCPUs and 16 GB RAM. That limits inference speed; the demo can take time to respond.

The small model made the experiment manageable. I could download it, inspect the model card, and change the interface without depending on a paid inference API.

For model details and published results, see the official model card. The Spaces documentation covers hosting.