← sources

Need maybe say "Use llama.cpp"

source: reddit (reddit:LocalLLaMA)published 2026-10-04 09:37ingested 2026-10-05 02:41

retained text

So I tried that miracle engine everyone is talking about.

Asked the IQ3\_S model to express its opinion on a post from this sub to measure the tps on a long-ish generation:

>Can you help with the following problem?  
>  
>So Kimi K2 is outdated, and so is GPT OSS 120b. Which of the modern open weights models can boast the least sycophancy? I need this both for creative/research assistant usage (sycophancy led me down blind alleys of my own bad ideas many times) and agentic coding (more sycophancy less bug noticing).

The thinking trace:

    We need answer user's question. Need likely provide current landscape as of 2026? We have get_datetime tool. Need know current date 2026? System says current date 2026-06-22. Need maybe use get_datetime? Could call to confirm. User asks about modern open weights models least sycophancy. Need likely discuss Kimi K2 outdated? 
    
    ...

10k tokens later it degrades to:

    Need maybe maybe include "Use 'for code, list constraints'."
    Need maybe maybe include "Use 'for code, list requirements'."

The same exact model in llama.cpp does produce a coherent answer without a doom loop.

document fetched 2026-10-05 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1wxbm8w/need_maybe_say_use_llamacpp