GLM-5.3-REAP37-MLX-4bit

MLX (Apple Silicon) build of GLM-5.3 — 744B-parameter glm_moe_dsa MoE (256 experts, top-8; MLA with DeepSeek-V3.2-style sparse attention) — quantized to 4-bit.

These files are modified: converted from the upstream bfloat16 release (GLM-5.3-BF16) to MLX and quantized; the architecture is unchanged. The multi-token-prediction layer (78) is not included.

Runtime — read this

This checkpoint bundles glm_moe_dsa.py (declared via model_file) and needs it:

pip install -U mlx-lm
mlx_lm.generate --model pipenetwork/GLM-5.3-REAP37-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-5.2/5.3 ship indexer weights on 21 (indexer_types: the other 57 "shared" layers reuse the previous full layer's top-k selection). A strict load of the release fails with 285 missing parameters; mlx_lm.load loads leniently and leaves those 57 indexers at random initialisation. Prompts up to 2048 tokens are unaffected (the indexer is bypassed below index_topk); beyond that, 57 layers attend to keys chosen by random projections. The bundled runtime implements the schedule as the reference does, plus the reference's 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, cached decode exact; strict loading of this checkpoint reports zero missing and zero unexpected tensors. Details and tests: https://github.com/PipeNetwork/glm53-mlx.

Size and what is quantized

0.0 GB on disk. RAM: 384 GB-class.

group	share of parameters	this build
routed experts (switch_mlp, 75 layers × 256)	724.8B (97.5%)	4-bit, group 64
attention (MLA), shared experts, dense layers 0–2, embeddings, lm_head	18.4B (2.5%)	4-bit, group 64
lightning indexer (21 layers), MoE router + correction bias, norms	0.3B	as stored (bf16 / fp32)

Source precision: the FP8 release is a lossy derivative of the bf16 one (dequantized FP8 weights differ from bf16 by up to 1.6e-2 on values of 0.46 — half an e4m3 step). The ladder row fp8 is the FP8 release itself measured against bf16: its error is the floor any FP8-sourced build inherits.

REAP pruning

This build keeps 161 of 256 routed experts per MoE layer (37% pruned; the 3 dense layers, attention, shared experts and the router are untouched), chosen by REAP saliency — mean router_weight × ‖expert_output‖ over 65,536 calibration tokens (wikitext-2 train, ten languages of Wikipedia and code; checked for zero 32-gram overlap with the eval set). Kept experts carry 78.3% of the layers' saliency mass on average. Ranking the experts on two disjoint halves of the calibration set picks the same kept set 89.8% of the time. The pruning was applied to the already-quantized 4-bit build, which is exactly equivalent to pruning bf16 and requantizing (expert subsetting and affine groups are on different axes). Saliency retention is not a quality measure — the perplexity below is.

Quality

Two measurements, because at 744B most of the ladder cannot be loaded on a 512 GB machine:

Per-layer divergence vs bf16 (scripts/eval_ladder.py): every decoder layer run in bf16 and in each recipe on identical inputs (16,384 tokens of wikitext-2), teacher-forced (each layer sees bf16 inputs — isolates its own damage) and free-running (each recipe feeds itself — what inference does). Relative L2 error of the layer output; lower is better.

recipe	teacher-forced (mean over layers)	free-running (final layer)	cosine (final)
8bit	0.00685	0.13119	0.98945
6bit	0.01465	0.16736	0.98389
5bit	0.02651	0.22521	0.97272
4bit	0.05161	0.35740	0.93390
mixed-4_8bit	0.02524	0.24951	0.96710
mixed-3_6bit	0.05242	0.42380	0.90624
fp8	0.01741	0.17321	0.98320

Perplexity on wikitext-2 (test), 288,627 tokens in 141 windows of 2048, for the builds that fit this machine, scored on identical windows:

build	size	perplexity [95% CI]
4bit	418.6 GB	2.8636 [2.6681, 3.0714]
mixed-4_8bit	427.8 GB	2.7420 [2.5533, 2.9477]
mixed-3_6bit	332.6 GB	3.0338 [2.8366, 3.2386]
REAP25-4bit	316.6 GB	3.2872 [3.0703, 3.5184]
REAP37-4bit	267.2 GB	3.8517 [3.6212, 4.0937]
REAP50-4bit	214.7 GB	5.0295 [4.7571, 5.3137]

Recommendation. For a 512 GB Mac, mixed 4/8-bit (427.7 GB): perplexity 2.7420, a paired 4.3% better than uniform 4-bit (ratio 0.9575 [0.9537, 0.9612], better on 98.6% of windows) for 9 GB more — the 2.5% of non-expert weights are worth their 8 bits, as on every model we have measured. Uniform 4-bit (418.6 GB) is the fallback when those 9 GB matter. Mixed 3/6-bit (332.6 GB) is the 384 GB-class option, at a real cost: 3.0338, +5.9% over 4-bit and +10.6% over mixed 4/8 — it leads the ladder for the first ten layers and then 3-bit expert damage compounds. Among the builds that cannot be run here, the ladder puts 8-bit closest to bfloat16 (free-running error 0.131), then 6-bit (0.167); the upstream FP8 release scores 0.173, between 6-bit and 5-bit, which is why these are converted from the bf16 release. 5-bit (0.225) sits just above mixed 4/8 (0.250) at 100 GB more.

Greedy generation (a collapse detector, not a ranking) is coherent on every published build.

License

GLM-5.3 license, as the upstream model. Port code: https://github.com/PipeNetwork/glm53-mlx.