← sources

Qwen 3.8 on 3090 comparisons

source: reddit (reddit:LocalLLaMA)published 2026-08-31 17:56ingested 2026-09-01 02:39

retained text

I’m getting \~35 tok/s with Qwen3.8-27B on my RTX 3090.

Another setup gets \~65 tok/s on the same GPU.



Is mine slow?



Mine:

• Q5 weights  
• Q8 KV cache  
• 32K context  
• 4 parallel slots  
• MTP off  



The faster setup:

• Q4 weights  
• Q4 KV cache  
• 131K context  
• 1 slot  
• MTP on    
  
The trade-offs:  
Q5 vs Q4: Different quantization precision  
  
Q8 vs Q4 KV: cache precision versus context capacity  
  
4 slots vs 1: concurrency versus single-user speed  
  
MTP (Multi-token prediction) on vs off: speculative speed versus memory and verification work  
  
Is there anything I'm missing? Any other variables that I should consider? 

For context, I'm using Llama.cpp 

document fetched 2026-09-01 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1w3lg41/qwen_38_on_3090_comparisons