Qwen 3.8 27B with Deepseek Harness

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
1

Using Deepseek Harness with Qwen 3.8 27B. Every messages gets repeated in each “thinking” cycle and the Qwen model thinks “user repeats the question”..

I swap to DeepSeek cloud models, and that works fine. So there is some sort of weird interaction with Qwen 3.8 that I don’t see with DeepSeek v4.

Has anybody encountered this and found a work around ?

2

Bro what quant and inference engine are you using? This sounds like a chat template or runtime config issue