GLM-5.2 Locally: Does REAP Pruning Make a 754B Model Fit?

LA

By Lefi Abdelmonem

Author · AI Local Check · Published July 19, 2026

GLM-5.2 (Z.ai) is widely rated one of the strongest open-weight models of 2026 — MIT licensed, a 1-million-token context, and roughly 754 billion parameters. Naturally people ask whether they can run it at home. And because GLM ships in REAP-pruned versions, there's a second, more interesting question: can pruning shrink a frontier MoE enough to fit a normal machine?

We ran the numbers on the real GGUF files. The answer is nuanced — and useful.

What GLM-5.2 needs (spoiler: a lot)

Even at the most aggressive quantizations, GLM-5.2 is server territory:

QuantizationMemory to load
IQ1_S (~1-bit)~203 GB
IQ1_M~214 GB
IQ2_M (~2-bit)~223 GB
IQ3_XXS~263 GB

For reference, Tencent's Hy3 drops to ~87 GB at 1-bit and does fit a 128 GB machine. GLM-5.2 does not — it's more than twice that.

What is REAP pruning?

REAP is expert pruning for Mixture-of-Experts models. An MoE like GLM-5.2 contains many "expert" sub-networks, but only a few fire per token. Pruning removes the least-used experts, producing a smaller model that keeps most of the behaviour. GLM-5.2 ships in community REAP variants at roughly 504B (about a third removed) and 380B (about half removed).

Does pruning make GLM-5.2 runnable at home?

This is the part people get wrong. Here's what the pruned builds actually need:

BuildParametersSmallest memoryFits a 128 GB machine?
GLM-5.2 (full)~754B~203 GBNo
GLM-5.2 REAP-504B~501B~197 GBNo
GLM-5.2 REAP50~381B~170 GBNo

Cutting half the experts only takes it from ~203 GB to ~170 GB. That's a real saving, but nowhere near enough to reach a 128 GB — let alone a 24 GB — machine. Pruning a frontier MoE does not turn it into a desktop model.

The takeaway: REAP is useful for cutting server costs, not for bringing a 754B model to your gaming PC. If a model needs 200 GB, removing a third of its experts still leaves ~197 GB.

Where pruning DOES pay off: the small end

The same technique is far more interesting on models that are already close to consumer size. GLM's smaller Flash line fits a normal GPU — and its REAP variant fits even better:

ModelParametersQuantizationVRAMRuns on
GLM-4.7-Flash-REAP-23B23BQ4_K_M~14.0 GB16 GB GPU ✅
GLM-4.7-Flash30BQ3_K_M~14.5 GB16 GB GPU ✅
GLM-4.5-Air110BQ3_K_M~54 GBWorkstation

The REAP-23B build gives you a GLM model at Q4_K_M in about 14 GB — comfortably inside a 16 GB card, with room for context. That's where expert pruning genuinely helps a home user.

So what should you actually run?

If you want GLM on your own hardware today, start with GLM-4.7-Flash or its REAP-23B build. If you have more VRAM, step up in size — see the best LLM for your VRAM, and GGUF quantization explained for what these labels mean. To check any model against your exact setup, use the calculator or pick your GPU.

The bottom line

GLM-5.2 is an excellent open model that stays in the data centre: ~203 GB at 1-bit, and still ~170 GB with half its experts pruned. REAP doesn't rescue frontier models for home use — but it does make already-small models smaller, and GLM-4.7-Flash-REAP-23B runs on a 16 GB GPU. Check what your PC can run.