As The Geek Learns AI Engineering Digestfrom As The Geek Learns

← Back to this issue

Monday, August 24, 2026

Try this on a Mac 2 stories

SenseNova-U1.5 8-bit T2I on MLX-Swift

Why now: four mlx-community artifacts were created this morning. From the 8-bit card: clone xocialize/sensenova-u1-swift, swift build -c release, hf download mlx-community/SenseNova-U1.5-8B-MoT-8bit, then sensenova-cli --weights … --prompt … --width 1024 --height 1024 --steps 50 --cfg 4.0. Card hardware hint: M5 Max, peak 22.9 GB for the 8-bit / 8-step path they quote. Distilled 8-step 4-bit/8-bit repos exist if 50-step is too slow. License Apache-2.0 on the conversion card. Do not treat the GitHub listing “3.2 s / 14.8 GB” and the card “7.4 s / 22.9 GB” as the same measurement.

8-bit cardSwift runtime

llama.cpp ≥ b10603 + GLM-4.5-Air MTP

Why now: MTP for this family is in a tagged macOS arm64 build for the first time this weekend. Published speedups are 4×3090 and Strix Halo Vulkan/ROCm, not Metal. Q4_K_M Air is on the order of tens of GB (HF GLM-4.5V Q4_K_M listed 66.9 GB); this is a high-unified-memory Mac, not a 24 GB laptop. Start from the macOS arm64 asset on b10603 or newer, plus a GLM-4.5-Air GGUF, or ggml-org/GLM-4.5V-GGUF. PR example flags: --spec-type draft-mtp --spec-draft-n-max 1 or 2.

b10603PRGLM-4.5V-GGUF