--trust-remote-codeThis repository now bundles glm_moe_dsa.py (declared via model_file in config.json), a fixed runtime
for this architecture, and needs it:
mlx_lm.generate --model pipenetwork/GLM-5.2-REAP25-MLX-4bit --trust-remote-code --prompt "..." --max-tokens 300
mlx-lm's own glm_moe_dsa builds a lightning indexer on all 78 layers, but GLM-5.2 ships indexer weights
on 21 (indexer_types: the other 57 "shared" layers reuse the previous full layer's top-k selection).
mlx_lm.load loads leniently and left those 57 indexers at random initialisation. Prompts up to 2048
tokens were unaffected (the indexer is bypassed below index_topk); beyond that, 57 of 78 layers attended
to keys chosen by random projections. The bundled runtime implements the schedule as the reference does
(plus fp32 indexer scores and router logits and the indexer LayerNorm epsilon); tiny-config parity against
transformers 5.16 is 4e-7 with the sparse path live, and a strict load of this checkpoint reports zero
missing and zero unexpected tensors. Details, tests and the GLM-5.3 builds made with it:
github.com/PipeNetwork/glm53-mlx. The weights are unchanged.
REAP expert-pruned + 4-bit MLX conversion of zai-org/GLM-5.2. Keeps the 192 most-salient experts per layer (of 256) → ~572B params, smaller/faster than the full model.
Pruned with REAP (Router-weighted Expert Activation Pruning, Cerebras / ICLR 2026): per MoE layer, experts are scored by mean(router_gate_weight × ‖expert_output‖) over a calibration set; the lowest-saliency experts are dropped and the router is sliced to the survivors. No retraining. n_routed_experts reduced 256→192.
| Variant | Experts | ~Params | Held-out PPL | vs full |
|---|---|---|---|---|
| full GLM-5.2 (4-bit) | 256 | ~750B | 1.447 | — |
| REAP25 (this repo) | 192 | ~572B | 1.481 | +2.3% |
| REAP37 | 160 | ~480B | 1.553 | +7.3% |
| REAP50 | 128 | ~394B | 1.990 | +37.5% |
This variant: PPL 1.481 (+2.3% vs full) — near-lossless. (Absolute PPL is low because the eval text is highly predictable; treat the numbers as relative degradation.)
Calibrated on the 4-bit GLM-5.2 (192 seqs × 1024 tok, prose + code); pruned during MLX conversion (no intermediate bf16). Requires the glm_moe_dsa / deepseek_v32 MLX path with per-layer indexer handling.
pip install mlx-lm
python -m mlx_lm generate --model pipenetwork/GLM-5.2-REAP25-MLX-4bit --prompt "Hello" -m 256
MIT (inherited from GLM-5.2). Quantization: {"group_size": 64, "bits": 4, "mode": "affine"}.
model-00001-of-00075.safetensors 5.3 GBmodel-00075-of-00075.safetensors 4.6 GBmodel-00002-of-00075.safetensors 4.4 GBmodel-00047-of-00075.safetensors 4.2 GBmodel-00051-of-00075.safetensors 4.2 GBmodel-00011-of-00075.safetensors 4.2 GBmodel-00043-of-00075.safetensors 4.2 GBmodel-00015-of-00075.safetensors 4.2 GBmodel-00055-of-00075.safetensors 4.2 GBmodel-00007-of-00075.safetensors 4.2 GBmodel-00027-of-00075.safetensors 4.2 GBmodel-00019-of-00075.safetensors 4.2 GBmodel-00059-of-00075.safetensors 4.2 GBmodel-00071-of-00075.safetensors 4.2 GBmodel-00023-of-00075.safetensors 4.2 GBmodel-00035-of-00075.safetensors 4.2 GBmodel-00067-of-00075.safetensors 4.2 GBmodel-00003-of-00075.safetensors 4.2 GBmodel-00063-of-00075.safetensors 4.2 GBmodel-00031-of-00075.safetensors 4.2 GBmodel-00039-of-00075.safetensors 4.2 GBmodel-00017-of-00075.safetensors 4.2 GBmodel-00046-of-00075.safetensors 4.2 GBmodel-00048-of-00075.safetensors 4.2 GBmodel-00010-of-00075.safetensors 4.2 GBmodel-00029-of-00075.safetensors 4.2 GBmodel-00038-of-00075.safetensors 4.2 GBmodel-00058-of-00075.safetensors 4.2 GBmodel-00072-of-00075.safetensors 4.2 GBmodel-00066-of-00075.safetensors 4.2 GBmodel-00060-of-00075.safetensors 4.2 GBmodel-00021-of-00075.safetensors 4.2 GBmodel-00050-of-00075.safetensors 4.2 GBmodel-00033-of-00075.safetensors 4.2 GBmodel-00052-of-00075.safetensors 4.2 GBmodel-00053-of-00075.safetensors 4.2 GBmodel-00062-of-00075.safetensors 4.2 GBmodel-00025-of-00075.safetensors 4.2 GBmodel-00026-of-00075.safetensors 4.2 GBmodel-00042-of-00075.safetensors 4.2 GBmodel-00070-of-00075.safetensors 4.2 GBmodel-00068-of-00075.safetensors 4.2 GBmodel-00069-of-00075.safetensors 4.2 GBmodel-00013-of-00075.safetensors 4.2 GBmodel-00040-of-00075.safetensors 4.2 GBmodel-00054-of-00075.safetensors 4.2 GBmodel-00057-of-00075.safetensors 4.2 GBmodel-00036-of-00075.safetensors 4.2 GBmodel-00009-of-00075.safetensors 4.2 GBmodel-00022-of-00075.safetensors 4.2 GBmodel-00049-of-00075.safetensors 4.2 GBmodel-00024-of-00075.safetensors 4.2 GBmodel-00012-of-00075.safetensors 4.2 GBmodel-00041-of-00075.safetensors 4.2 GBmodel-00064-of-00075.safetensors 4.2 GBmodel-00014-of-00075.safetensors 4.2 GBmodel-00018-of-00075.safetensors 4.2 GBmodel-00032-of-00075.safetensors 4.2 GBmodel-00061-of-00075.safetensors 4.2 GBmodel-00065-of-00075.safetensors 4.2 GBmodel-00030-of-00075.safetensors 4.2 GBmodel-00056-of-00075.safetensors 4.2 GBmodel-00074-of-00075.safetensors 4.2 GBmodel-00028-of-00075.safetensors 4.2 GBmodel-00034-of-00075.safetensors 4.2 GBmodel-00037-of-00075.safetensors 4.2 GBmodel-00006-of-00075.safetensors 4.2 GBmodel-00020-of-00075.safetensors 4.2 GBmodel-00044-of-00075.safetensors 4.2 GBmodel-00016-of-00075.safetensors 4.2 GBmodel-00045-of-00075.safetensors 4.2 GBmodel-00073-of-00075.safetensors 4.2 GBmodel-00005-of-00075.safetensors 4.2 GBmodel-00008-of-00075.safetensors 4.2 GBmodel-00004-of-00075.safetensors 4.2 GBHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo pipenetwork/GLM-5.2-REAP25-MLX-4bit ./model-folderzai-org/GLM-5.2