antirez: for local LLM inference, fast prefill beats high t/s — 20-30 t/s is enough
teortaxesTex · x · 2026-09-29
antirez argues tokens-per-second is the wrong headline metric for local inference: the real game is shortening the thinking phase, and 20-30 t/s suffices with fast prefill — a hint Chinese open-model providers may act on. teortaxesTex counters that providers know this already; generating training data and training simply takes far more compute.
More from Infra
- Gemini 3.5 Flash now runs on air-gapped Google Distributed Cloud, offline up to 12 months — shashib · 2026-11-16
- Gemma.c: full Gemma 1/2/3 multimodal inference in a single C file, ~72 tok/s on CPU at INT8 — This-Purchase-3325 · 2026-11-16
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16