antirez: stop chasing tokens/s — fast prefill and short thinking beat 20-30 t/s myths
antirez · x · 2026-09-29
antirez argues that tokens per second is the wrong primary metric for local LLM inference: Chinese open-model providers will soon compete on making the thinking phase as short as possible, and 20/30 t/s is enough with fast prefill. He adds that upcoming models will make DGX Spark and Strix Halo far more compelling, and that local hardware shouldn't be evaluated in a vacuum — its value depends on the class and strengths of models being released.
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