GLM-5.3
Quant Diff
CrossingBits re-quant of official zai-org/GLM-5.3 provider block-FP8 (e4m3, 128×128).
| Recipe | weight_fp4_activation_fp8, group 32, ue8m0 / e4m3 |
| Routed experts + dense FFN + most attn | FP4 E2M1 weights + FP8 E4M3 activations |
mlp.shared_experts.{gate,up,down}_proj |
MXFP8 W8A8 g32 (MegaMoE) |
| BF16 | router (mlp.gate), MLA absorb (kv_b / k_b / v_b), indexer.weights_proj, embed, lm_head, shared_head |
| MTP | enabled (layers.78) |