GLM-5.2-NVFP4-NF3-Hybrid
This was the original hybrid (fp8 non-expert tier). The successor is a drop-in upgrade with
the same quality and much more context:
- Same intelligence — GPQA-Diamond 88.38 (parity with the full model), byte-identical
NF3 3-bit + NVFP4 4-bit experts.
- 2.44× the KV pool — the non-expert tier is now native MXFP8 (served without a
dequant-on-load penalty), freeing VRAM for ~126k → ~267k+ context at DCP4.
- Ready-to-run — a
docker-compose.yml + the :v2 image are baked and public.
There's no reason to use this older version.
The full GLM-5.2 — 753B, all 256 experts, no pruning — on 4× 96GB GPUs, via NF3 (our
in-house 3-bit expert format) and a custom intra-layer expert-by-expert mixed-precision
sm120 kernel: each expert quantized to its own bit-width by saliency, within a single MoE
layer. By madeby561.
