Open weights · Published checkpoints

Models built for on-prem inference.

Canonical model names; precision and packaging identified separately.

Published models

NF4 Lloyd Experts · GPTQ

Lloyd-codebook NF4 routed experts quantized with activation-aware GPTQ and error feedback.

Hybrid FP8 · NVFP4 · MXFP4

Mixed-precision checkpoint built for memory-constrained, multi-node inference on four DGX Spark systems.

MXFP4 Experts · GPTQ

Activation-calibrated MXFP4 expert weights for the GLM-5.2 mixture-of-experts architecture.

INT8 Dynamic

Dynamically quantized text-generation checkpoint of DiffusionGemma 26B-A4B-it.