@brianbellx: I removed 423 GB from GLM‑5.2 without changing the model. 1,403 GB → 980 GB. 753B weights. Bit for bit exact. No quanti…
Summary
A technique to remove 423 GB from GLM-5.2 (753B weights) without quantization or retraining, achieving bit-exact compression by keeping weights compressed in VRAM.
View Cached Full Text
Cached at: 07/13/26, 07:51 AM
I removed 423 GB from GLM‑5.2 without changing the model.
1,403 GB → 980 GB. 753B weights. Bit for bit exact. No quantization or retraining.
The weights remain compressed in VRAM instead of rebuilding the full model first.
Full writeup and repo in the next post.
yes, but quantizing would shave off all savings. so wouldn’t even make a difference.
im not sure what the floor is. still working towards finding it out.
much appreciated!
yea, theres more work to be done. i don’t think anyone knows what the floor for this is.
Similar Articles
jlnsrk/GLM-5.2-colibri-int4
Pre-converted int4 quantized weights for the GLM-5.2 744B MoE model, designed to run on consumer hardware with ~25 GB RAM using the colibrì engine.
@UnslothAI: GLM-5.2 can now be run locally! The 2-bit model retains ~82% accuracy after we shrunk it from 1.51TB to 238GB (-84% siz…
UnslothAI announces GLM-5.2, Z.ai's strongest open model with 744B parameters, now runnable locally via dynamic GGUF quantization reducing size by ~84% to 239GB while retaining ~82% accuracy. It fits on 256GB Macs and supports long-context, reasoning, and agentic tasks.
@antirez: GLM 5.2, Q2_K routed experts (effectively ~2.6 bits) running with SSD streaming on an M5 Max 128GB computer.
GLM 5.2 model runs with Q2_K quantized routed experts (effective 2.6 bits) using SSD streaming on an M5 Max 128GB computer.
@0xSero: We found a way to run GLM-5.2 with full context in vLLM without pruning. - top 32 experts NVFP4 - rest fp3 - intel auto…
A community researcher enabled running GLM-5.2 (753B parameters, all 256 experts) in vLLM without pruning via a hybrid quantization (NVFP4, NF3, MXFP8), fitting on 4×96GB GPUs with ~307k KV cache and near-FP8 accuracy.
@tolak_eth: I wanted to share how we avoided spending roughly $160k/year to host GLM-5.2 with its full 1M context. When GLM-5.2 lau…
Phala avoided $160k/year hosting costs for GLM-5.2 with full 1M context by quantizing MoE experts to 4-bit and keeping critical parts in FP8/BF16, achieving the same benchmark results on a single 8×H200 node and releasing the optimized model GLM-5.2-W4AFP8 on Hugging Face.