Checkpoints ready for your own hardware.
Canonical model names; precision and packaging identified separately. Each one is paired with a recipe that ran it.
Published checkpoints
Hugging Face profile ↗| No. | Checkpoint | Package | Tags | |
|---|---|---|---|---|
| 01 | GLM-5.2FP8 attention, NVFP4 dense/shared experts and MXFP4 routed experts for 4× DGX Spark. | Hybrid FP8 · NVFP4 · MXFP4 | 425B · Mixed precision · DGX Spark | ↗ |
| 02 | GLM-5.2Activation-calibrated MXFP4 experts for GLM-5.2. | MXFP4 Experts · GPTQ | MXFP4 · GPTQ · Experts | ↗ |
| 03 | DiffusionGemma 26B-A4B-itW8A8 INT8 dynamic quantization of DiffusionGemma 26B-A4B-it for Ampere GPUs. | INT8 Dynamic | 26B · INT8 · Text generation | ↗ |
Quantizations made here, published as they were measured. Precision is stated per component where it differs; nothing is labelled a single precision when it isn't.