GLM-5.3-SF-WInt4AInt8-ASC
SiliconFlow / crossingbits quantization of ZhipuAI/GLM-5.3 (zai-org/GLM-5.3).
This is the Ascend W4A8 layout (act_group_size=128, backend_layout_mode=layout_only). Source W4 qweight tensors are not stored in the main checkpoint; routed-expert weights live in backend_layouts-ascend_w4a8-*.safetensors.
Quantization
| Tool | crossingbits |
| Recipe | weight_clip (channel-wise, group_size=-1, symm_q=true) |
| Activations | INT8, G128 |
| Save | save_backend_layouts=[ascend_w4a8], layout_only |
| MTP | yes (use_mtp=true) |
Packing
| Module | dtype |
|---|---|
routed mlp.experts.*.{gate,up,down}_proj |
W4A8 (Ascend sidecar) |
L0–L2 dense MLP, shared experts, attn q_b / o_proj, indexer wq_b / wk, MTP eh_proj |
W8A8 |
router mlp.gate, MLA q_a / kv_a / kv_b / k_b / v_b, indexer.weights_proj, shared_head, embed, lm_head |
BF16 |
Base model
GLM-5.3 uses the same base as GLM-5.2; the gains are from post-training. See the GLM-5.3 blog and ZhipuAI/GLM-5.3.
License
Same as the base checkpoint (LICENSE in this repo). Not an MIT/Apache relicense.