GLM-5.3 MLX - a pipenetwork Collection

pipenetwork 's Collections

GLM-5.3-Flash MLX

Qwen3.8-Flash-Next MLX

Qwen3.8-2.4T-A95B MLX (REAP)

Muse-Glimmer 30B MLX

Inkling-Small · MLX

DeepSeek-V4-Flash · MLX

Nemotron TwoTower · MLX

Qwen3.6-35B-A3B — MLX

Ornith-1.0-397B — MLX

GLM-5.2 REAP (expert-pruned)

Rio-3.1-Open-30B MLX

Gemma-4-26B-A4B-it MLX

Gemma-4-31B-it MLX

Kimi-K2.7-Code MLX

Holo-3.1 MLX (computer-use)

Frog (SWE/debugging) MLX

Nemotron-3 MLX (Apple Silicon)

GLM-5.3 MLX

updated about 18 hours ago

Apple Silicon (MLX) builds of GLM-5.3 (744B) with a fixed glm_moe_dsa runtime: github.com/PipeNetwork/glm53-mlx

pipenetwork/GLM-5.3-MLX-mixed-4_8bit

Text Generation • 119B • Updated about 18 hours ago • 80 • 1

Note 427.7 GB, 512 GB Mac. ppl 2.7420 — the build to use: 4.3% better than uniform 4-bit for 9 GB more.
pipenetwork/GLM-5.3-MLX-4bit

Text Generation • 116B • Updated about 18 hours ago • 47 • 1

Note 418.6 GB, 512 GB Mac (tight). ppl 2.8636.
pipenetwork/GLM-5.3-MLX-mixed-3_6bit

Text Generation • 95B • Updated about 18 hours ago • 71

Note 332.6 GB, 384 GB-class. ppl 3.0338 (+5.9% vs 4-bit): 3-bit experts compound over 78 layers.
pipenetwork/GLM-5.3-MLX-8bit

Text Generation • 209B • Updated about 18 hours ago • 3

Note ~800 GB, two machines. Closest to bf16 on the per-layer ladder (free-running error 0.131, cos 0.989).
pipenetwork/GLM-5.3-MLX-6bit

Text Generation • 163B • Updated about 18 hours ago • 14

Note 604 GB, two machines. Ladder 0.167 — better than the upstream FP8 release (0.173).
pipenetwork/GLM-5.3-MLX-5bit

Text Generation • 140B • Updated about 18 hours ago • 80

Note 512 GB, two machines. Ladder 0.225.
pipenetwork/GLM-5.3-REAP25-MLX-4bit

Text Generation • 88B • Updated about 18 hours ago • 7

Note 316.6 GB. 192 of 256 experts per layer. ppl 3.2872 (+14.8% vs 4-bit).
pipenetwork/GLM-5.3-REAP37-MLX-4bit

Text Generation • 74B • Updated about 18 hours ago • 7

Note 267.2 GB. 161 of 256 experts. ppl 3.8517 (+34.5% vs 4-bit).
pipenetwork/GLM-5.3-REAP50-MLX-4bit

Text Generation • 60B • Updated about 18 hours ago

Note 214.6 GB, 256 GB Mac. 128 of 256 experts. ppl 5.0295 (+75.6% vs 4-bit) — coherent, but GLM-5.3 tolerates pruning poorly; the 512 GB builds are a different class.