wrldsuksgo2mars/GLM-5.3-EXL3-K4-v1

GLM-5.3 EXL3 K4

This is a calibrated 4-bit EXL3 quantization of zai-org/GLM-5.3.

The routed expert gate_proj, up_proj, and down_proj tensors in base-model layers 3 through 77 use the EXL3 K4 representation. Other tensors, including the native MTP overlay, retain their source representation. The repository contains 46 safetensors shards and the complete quantization metadata in quantize_config.json.

Compatibility

This is an EXL3 checkpoint, not a drop-in GPTQ or standard Transformers checkpoint. Use a serving runtime that supports GLM-5.3 MoE/DSA and this EXL3 tensor layout. Check runtime compatibility before downloading the full model.

Tool-use result

The best observed tool-eval-bench run scored 91/100 (125/138 raw points) with fixed DFlash2 draft width 5. This is a single local evaluation result and is not an official upstream GLM-5.3 benchmark.

License

This derivative checkpoint is distributed under the GLM-5.3 model license. Review and comply with the license terms of the original zai-org/GLM-5.3 model before use.