Do we really need all these "speed" posts?
retained text
I appreciate the research and development that is going into these open source tools. However, there appears to be multiple posts on a daily basis about reaching 300 t/s, and then reading into it: 1. Running on Q1\_XXXS REAP quant 2. Greedy decoding 3. TAIL\_SKIP (I only have 5GB RAM, but if the API ignores all experts that are not in RAM, then I can still say I generated something and call a speed record. YAY! Without any KLD benchmark data being presented, are these posts really of value to this community? What are your thoughts?
document fetched 2026-11-03 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xljjf2/do_we_really_need_all_these_speed_posts