GLM-5.3-Flash-REAP50-NVFP4 (v1, superseded; do not use)

The first published NVFP4 build of a 50% expert-pruned zai-org/GLM-5.3-Flash. It is kept only as a record. It carries a known healing defect, and it has never been evaluated or served.

Use GLM-5.3-Flash-REAP50-NVFP4-v2 instead. That repo holds 98.2 GiB of weights with top-1 agreement of 0.839 against the unpruned model, and they have been served on Jetson AGX Thor.

The known defect

This build was quantised from the v1 FP8 master before that master's healing gain was corrected. Every retained expert's output is therefore scaled by a first-moment gain (median 0.69) where the measured value is 0.9111. That makes the MoE pathway about 0.76× too small relative to attention, the shared experts and the residual stream, in all 42 MoE layers. The FP8 master was fixed on 2026-08-28 by rescaling its block scales. This repo was never corrected; the corrected NVFP4 build is v2.

Model overview

Base model	zai-org/GLM-5.3-Flash: MIT, 321.3B parameters by safetensors count, 305.8 GiB in FP8
Architecture	Glm5NextForConditionalGeneration (glm5_next). 45 decoder layers: 34 KDA linear-attention layers and 11 MLA layers with DSA sparse attention. Layers 0–2 are dense MLP, layers 3–44 are MoE. Hidden size 4096. Includes a 24-block vision tower.
Parameters	161.7B after pruning. This is the element count of the v1 FP8 master, which has the same tensors; the Hub's count for this repo counts packed bytes, not weights.
Routed experts	144 of 288 kept in every MoE layer. Top-8 routing and 1 shared expert, both unchanged.
MTP block	Not included. There are no layer-45 tensors, and num_nextn_predict_layers is 0.
Context length	1,048,576 (max_position_embeddings, unchanged)
Precision	Routed and shared experts are NVFP4: weight-only, group size 16, FP8 E4M3 group scales and an F32 global scale. Everything else is BF16: KDA and MLA attention, the DSA indexer, dense layers 0–2, routers, embeddings, lm_head, the vision tower and the hyper-connection weights. There are no input_scale tensors.
Format	compressed-tensors, nvfp4-pack-quantized
Size on disk	95.9 GiB (103,022,994,376 bytes, 62 safetensors shards)
Licence	MIT, inherited from the base model

Which version should I use

Weights	Location	Selection	Healing	Size	Top-1 vs unpruned	Use
v2 NVFP4	GLM-5.3-Flash-REAP50-NVFP4-v2, root	pass-2 REAP	measured per-layer scalar	98.2 GiB	0.8386	Recommended for 128 GB devices
v2 FP8	GLM-5.3-Flash-REAP50-FP8-v2, root	pass-2 REAP	measured per-layer scalar	160.6 GiB	0.8425	Recommended master, when memory allows
v3 NVFP4	GLM-5.3-Flash-REAP50-NVFP4-v2, pass3-hope/	HOPE, protect_frac 0.04	first-moment fallback	93.6 GiB	not evaluated	Experimental
v3 FP8	GLM-5.3-Flash-REAP50-FP8-v2, pass3-hope/	HOPE, protect_frac 0.04	first-moment fallback	160.6 GiB	not evaluated	Experimental
v1 FP8	GLM-5.3-Flash-REAP50-FP8	pass-1 REAP	measured scalar (re-fitted 2026-08-28)	157.0 GiB	0.8370	Superseded; no MTP block
v1 NVFP4	this repo	pass-1 REAP	superseded first-moment gain	95.9 GiB	not evaluated	Superseded; do not use

GGUF builds for llama.cpp, made from the v2 FP8 weights, are in GLM-5.3-Flash-REAP50-GGUF; see that card for their details.

Quickstart

There is none. This build has not been loaded by any runtime, and it is not supported by the author's Thor CUDA server, which supports only the v2 NVFP4 weights. The repo also has no preprocessor_config.json or processor_config.json. Use GLM-5.3-Flash-REAP50-NVFP4-v2.

Evaluation

Not evaluated. No teacher-forced or generative evaluation has been run on these NVFP4 weights. The v1 FP8 master (0.8370) was evaluated only after its healing re-fit, so that result does not describe this build, which carries the pre-fix gain.

Hardware and memory

The build machine is an NVIDIA Jetson AGX Thor (128 GB unified memory, sm_110a). These weights have not been run on it or on any other hardware.

How it was made

Same selection as GLM-5.3-Flash-REAP50-FP8 (v1). That is REAP (arXiv:2510.13999) on 256 calibration samples, about 0.52M tokens, keeping 144 of 288 experts in each of the 42 MoE decoder layers, with the MTP block left out. Routed and shared expert weights were then converted to weight-only NVFP4 from the FP8 master, still carrying its first-moment healing gain. The FP8 MLA projections and dense layers 0–2 were dequantised to BF16.

Limitations

• Healing defect (see above): the MoE output is mis-scaled in every MoE layer.
• Unevaluated and never run.
• No MTP block.
• Superseded by v2 on every measured axis.

Changelog and corrections

• 2026-08-28: published. Later that day the card was flagged: this build carries the superseded healing gain, and the FP8 master was corrected instead.
• 2026-11-06: card rewritten and verified against repo files and build artifacts; a 2026-10-26 automated rewrite contained errors.

License and attribution

Released under the MIT licence, the base model's licence (see LICENSE). The base model is GLM-5.3-Flash by Z.ai; all credit for the architecture and pre-training belongs to its authors. The pruning criterion is REAP: Lasby et al., REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression, arXiv:2510.13999.

Citation

@misc{devaney2026glm53reap50,
  title        = {GLM-5.3-Flash-REAP50: 50\% expert-pruned GLM-5.3-Flash in FP8 and NVFP4},
  author       = {Patrick Devaney},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/patrickbdevaney/GLM-5.3-Flash-REAP50-NVFP4}}
}