HuggingChat should use the original Deepseek R1 instead of the Distilled Qwen 32 B parameter version.

[dagoatzmcclusky]

Bug description

The point is, when I use the distilled 32B Qwen version of deepseek, it significantly hallucinated like gpt-4o in many mid-tier tasks. Instead, when I switched back to the original R1 model on HuggingFace, the model ddid not hallucinate.

Steps to reproduce

Maybe the devs should replace deepseek r1-distill-32b-qwen model with the original deepseek-r1 model. The latter is proved to be more capable at handling more complex tasks and does not hallucinate like the r1-distill-32b-qwen model.
Please fix this problem as soon as possible, the current deepseek r1-distilled-32b model is now unusable after this type of error.

Screenshots

The model hallucinates every time i prompt it to write a vietnamese essay.

Specs

• OS: Ubuntu 24.04 LTS
• Browser: Vivaldi

Notes

Since the current deepseek-r1-distill-32b model is based on the qwq-32b-preview model in which the latter is already available on huggingchat, so why having deepseek-r1-distill-32b model anyway? It's best to replace it with the original deepseek-r1.

[julien-c]

we are working on it!