Lossless BF16 Weight Compression for GLM-5.2
Lossless BF16 Weight Compression for GLM-5.2
Source: https://x.com/i/status/2076432307687215494
📌 In mid-July 2026, Brian Bell reported cutting ~423 GB from open-weight GLM-5.2 by lossless packing of BF16 values—shrinking storage from ~1,403 GB to ~980 GB with bit-for-bit reconstruction and no quantization or retraining. The approach exploits the real distribution of trained weights so models can stay compressed in VRAM without materializing a full uncompressed copy.
📦 What was compressed
Zhipu AI’s GLM-5.2 open BF16 MoE weights (~744–753B total params, ~40B active per token). Full scan covered 282 shards and 59,509 BF16 tensors—normally well over a terabyte on disk.
🔬 How it works
Not quantization or fine-tuning: lossless packing of how trained BF16 values are distributed. A byte-split path recovers the high byte from a codebook, indices, and an escape stream for rare symbols; the low (mantissa) byte stays raw.
📉 Measured savings
K15 charged-format accounting: ~30.17% smaller (~1,403.19 → 979.87 GiB). Byte-split decode costs ~12.005 bits/weight (~24.97% under native BF16) and matched all tensors bit for bit.
⚡ Why it matters
Targets residency and disk/VRAM cost for self-hosting and multi-GPU serving without accuracy trade-offs. Runtime pitch: keep weights compressed in VRAM and use them without first expanding a full uncompressed copy.
🔗 Where to look
Public July 2026 experiment under brianbell-x/weight-compression, with a writeup and the original X post announcing the ~423 GB cut.
Key facts
| Fact | Value |
|---|---|
| Model | GLM-5.2 (zai-org), open BF16 MoE weights |
| Claimed size cut | 1,403 GB → 980 GB (~423 GB, ~30%) |
| Parameters cited | 753B BF16 weights (~40B active per token) |
| Fidelity | Bit-for-bit exact; no quantization or retraining |
| Decoded rate | ~12.005 bits/weight (~24.967% under BF16); all 59,509 tensors verified |
| Evidence scope | Full scan of 282 shards; K15 ~30.168% accounting; exact byte-split inverse check |
Details
GLM-5.2 is Zhipu AI’s (z.ai) large Mixture-of-Experts model—on the order of 744–753B total parameters with about 40B active per token. Its open BF16 weights normally need well over a terabyte of memory, a practical barrier for self-hosting and multi-GPU serving even when quality is left alone.
In mid-July 2026, developer Brian Bell (@brianbellx) reported cutting about 423 GB from the model without changing its parameters: storage went from roughly 1,403 GB to about 980 GB. The claim is not quantization or fine-tuning, but lossless packing that exploits how trained BF16 values are actually distributed, so every tensor can be reconstructed bit for bit.
On the full model (282 shards, 59,509 BF16 tensors), “K15 charged-format accounting” prices a layout at about 30.17% smaller (roughly 1,403.19 → 979.87 GiB). A separate byte-split path was round-tripped exactly: the high byte of each weight is recovered from a codebook, indices, and an escape stream for rare symbols; the low mantissa byte is kept raw. That decoded form costs about 12.005 bits per weight—about 24.97% under native BF16—and matched all tensors bit for bit.
The runtime pitch is that weights can stay compressed in VRAM and be used without first materializing a full uncompressed copy. Work is public under an ongoing July 2026 experiment in the brianbell-x/weight-compression repo.