Video gen model runs on a 16GB MacBook Air — a 20s clip takes 58 hours to render
DaLyon92x · reddit · 2026-11-10
A developer ran the h3 video generation model on an old 16GB M1 MacBook Air using his fork of h3.c — a pure C/Metal inference engine with no Python, adding int8 quantization and SSD streaming. The int8 checkpoint with a folded turbo LoRA needs 10GB RAM (13.5–14.6GB peak). Timings at 6 steps: 0.4MP/3.75s takes 1h39m, 0.9MP/3s takes 4h18m, and a 20-second 1280x704 clip chained across 9 links of 73 frames took 58 hours. Context chaining works — the demo includes a 20s talking head (5 links, 17h). Usable on a 48GB M5 Pro for background rendering; he asks owners of M6 32GB mini or M5 Ultra 96GB machines to test the fork.
More from Infra
- PagedAttention explained: managing KV cache like OS virtual memory — NielsRogge · 2026-11-10
- TSMC October revenue growth accelerates to +60% y/y on AI chip demand — firstadopter · 2026-11-10
- Archil claims self-hosted Spark on its S3-as-local-disk storage runs 2-5x faster than Databricks for under half the cost — massimosgrelli · 2026-11-10
- 467-run ComfyUI quant benchmark across 10 GPUs: int8 ConvRot most faithful, nvfp4 fastest but drifts — mybrianonacid · 2026-11-10
- llama.cpp Adds --shard Flag for Multi-Location Model Loading and Separate PLE Files — pmttyji · 2026-11-10
- Four must-haves for agent sandbox providers, from hands-on setup — pramodk73 · 2026-11-10