Open-source h3metal runs image gen on M5/M6 matrix units, 2.1-2.3x faster than Draw Things
audenvenn · reddit · 2026-11-15
h3metal is an open-source (Apache-2.0) image-generation engine for Apple silicon that runs transformers as int8 matrix products on the matrix units Apple added to M5/M6 GPU cores. Benchmarks on an M6 Mac mini vs Draw Things CLI: FLUX.2 [klein] 4B in 7.0s vs 17.0s, Z-Image-Turbo in 17.1s vs 40.0s, with per-step denoising 2.1-2.3x faster. Techniques include Hadamard-rotated int8×int8 products, flash attention with int8 scores, implicit-GEMM VAE convolutions, and LoRAs applied inside int8 products. Also runs FLUX.2 [klein] 9B and MiniMax-H3-Image (2048×2048 in 16s). Requires M5/M6-class GPU and macOS 26; pip-installable with one-command generation and a local web UI.
More from Infra
- ggml tensor library trends on GitHub with 15.8k stars, up 154 in a day — ggml-org · 2026-11-15
- speech.cpp v0.11.0 lands Haru-4B, a full-duplex voice chat model running locally in C++ — alexcovo_eth · 2026-11-15
- BiRefNet-lite runs fully on a phone NPU at 1024px in 23.7ms on Snapdragon 8 Elite Gen 5, open weights — subhashg547 · 2026-11-15
- Azure outage draws silence: big consultancies avoid vendor concentration risk talk — DavidLinthicum · 2026-11-15
- 2 Starship launches equal 20 Falcon 9 missions as 52 V3 sats match 500+ V2 capacity — XFreeze · 2026-11-15
- HPE Beats Estimates but Stock Slides as Analysts Doubt AI Server Margins — DavidLinthicum · 2026-11-15