0xSero/GLM-5-REAP-50pct-BF16-GGUF

Support this work: donate.sybilsolutions.ai

REAP surfaces: GLM | MiniMax | Qwen | Gemma | Paper | Code | PR17 | Cerebras Collection

GLM-5 REAP-50% BF16 GGUF

Expert-pruned GLM-5 (744B -> ~372B params, 256 -> 128 routed experts) in bf16 GGUF format for llama.cpp inference. This is the full-precision intermediate used to produce quantized GGUFs.

Model Details

Property	Value
Base model	zai-org/GLM-5 (744B, 256 routed experts)
Pruning	REAP saliency pruning, 50% expert removal (256 -> 128 experts)
Format	bf16 GGUF (full precision, no quantization loss)
Size	~711 GB
Architecture	GlmMoeDsaForCausalLM (MLA + Mixture of Experts + DSA indexer)
Context	202,752 tokens
Active params	~20B per token (8 of 128 experts selected)

Purpose

This bf16 GGUF serves as the source for all quantized variants. Use it to:

• Produce custom GGUF quantizations with llama-quantize (Q2_K, Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0, etc.)
• Generate importance-matrix (imatrix) calibrated quantizations for higher quality at low bit rates
• Run full-precision inference if you have sufficient VRAM/RAM (~711 GB)

Quantized Variants

Variant	BPW	Size	Repo
Q3_K_M	3.82	~170 GB	0xSero/GLM-5-REAP-50pct-Q3_K_M-GGUF

Usage

# Download
huggingface-cli download 0xSero/GLM-5-REAP-50pct-BF16-GGUF --local-dir GLM-5-REAP-50pct-BF16-GGUF

# Quantize to any format
llama-quantize GLM-5-REAP-50pct-BF16-GGUF/GLM-5-REAP-50pct-BF16.gguf output-Q4_K_M.gguf Q4_K_M

# With importance matrix for better quality
llama-quantize --imatrix imatrix.dat GLM-5-REAP-50pct-BF16-GGUF/GLM-5-REAP-50pct-BF16.gguf output-IQ3_M.gguf IQ3_M

Pipeline

1. REAP pruning of GLM-5 FP8 (256 -> 128 experts) using layerwise saliency scores
2. FP8 scale repair (block-quantized scale tensors corrected from 256-expert to 128-expert layout)
3. FP8 -> bf16 GGUF conversion via patched convert_hf_to_gguf.py (auto-dequants FP8, splits fused gate_up_proj, handles 3D expert tensors)

Why GGUF?

The original FP8 safetensors model has a known KV cache NaN bug in HuggingFace Transformers' GlmMoeDsaAttention implementation. llama.cpp bypasses this entirely with its own inference engine, producing correct output with working KV cache.

Source Models

• FP8 pruned model: 0xSero/GLM-5-REAP-50pct-FP8
• Original base model: zai-org/GLM-5