8B Model Hits 60 Tokens/sec on Phone CPU Alone, No GPU or NPU Needed
const_reborn · x · 2026-09-29
At the Exploit conference, @jondurbin reported running an 8B model entirely on a phone's CPU — no GPU or NPU — at roughly 60 tokens per second.
The test used real devices rented through Qualcomm Device Cloud, and notably was measured before any optimization work. That suggests meaningful headroom for on-device LLM performance once mobile-specific quantization and scheduling tweaks land.
More from Infra
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16
- Colossus 2's 1.2M B200s could generate ~$400B/year in tokens at DeepSeek Flash rates — bookwormengr · 2026-11-16
- BIS issues rule blocking remote access to US AI chips via third-country data centers — teortaxesTex · 2026-11-16
- Qdrant 1.20.1: 50% fewer cold-storage reads, ~2x faster quantization, agent-ready setup — qdrant_engine · 2026-11-16