What tok/s would you tolerate running local LLMs? 15 is the cutoff
challis88ocarina · reddit · 2026-11-01
A Reddit thread asks what token/s throughput people can live with when running models like Fable/Carbon/Astra locally. The OP's cutoff is 15 t/s; he notes Muse Spark drops below 10 t/s at 300k context but is still faster than DSv41 in wall-clock terms thanks to token efficiency.
More from Models
- Mysterious New Model Shows GPT-4o-Style 'Hallucinatory' Outputs, Reviving #keep4o Debate — aiamblichus · 2026-11-01
- User praises 'opus-5.5' for Minecraft builds oozing taste — unverified model — adonis_singh · 2026-11-01
- OpenAI to merge ChatGPT and Codex usage limits on Plus, Pro and Business plans in December — koltregaskes · 2026-11-01
- Comparing GPT generations on alt-history scenarios: gpt-6 is 'another level' — prd_008 · 2026-11-01
- Debate: If local models get "good enough", OpenAI and Anthropic are toast — TheZachMueller · 2026-11-01
- Running a 125B Model on a $2,000 PC: Qwen 3.8 Flash Next with Strata Tested — RelevantEmergency707 · 2026-11-01