FlashLabs Ships GLM-5.3-MLX, Runs a 743B Model on a 512GB Mac
The full GLM-5.3, not a distilled or Flash variant, now runs locally on a single 512GB Mac, which is the step that moves a frontier open-weight model onto everyday hardware.
Reporting from 1 source: ASCII.jp.
FlashLabs released GLM-5.3-MLX, an Apple Silicon-optimized MLX quantization of Z.ai's open-weight GLM-5.3 (about 743B total parameters, 40B active, 1M-token context), on Hugging Face. The OrcaSAQ method cuts the FP8 size from about 1.51TB to roughly 322GB at 2-bit, a 79 percent reduction, while keeping about 80.2 percent accuracy. A 512GB Mac with M3 Ultra runs the 2-bit, 3-bit, and 4-bit builds locally.
The release extends an August 28 build: FlashLabs already shipped GLM-5.3-Flash, a 320B-parameter variant, in MLX form, and GLM-5.3-MLX is the full model. OrcaSAQ needs no calibration data and keeps accuracy-critical tensors at higher bit widths, which is how the 2-bit build holds about 80.2 percent of FP8 accuracy while shrinking the footprint by roughly 79 percent.
The four variants split by use case. A 512GB Mac with M3 Ultra runs the 2-bit and 3-bit builds comfortably and fits the 4-bit build. The 6-bit version, which reaches a cosine similarity of 0.9997 against FP8, targets two 512GB Macs linked through mlx.distributed or eight NVIDIA H200 GPUs. There is no 8-bit build because expert weights would grow larger than the FP8 source weights.
Unsloth also ships a GGUF version for llama.cpp environments. OrcaRouter's API serves the model at full precision for setups that cannot run it locally.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.