Model overview · GLM
GLM-5.3-Flash (320B-A18B)
Generated from Architecture IR 0.1.0 · revision f93128cf7f
· report a correction
Overview
Fact sheet
- Layers
- 45
- Hidden size
- 4096
- Attention heads
- 64
- Vocab
- 154,880
- Context
- 1M
- Params
- 320B total / 18B active
Architecture Genome
One column per layer — top: attention, bottom: FFN/MoE.
Layer
attention (GQA/MLA) linear / recurrent attention MoE dense FFN
Evidence
Every visible number resolves to one of these claims. Inferred or unknown values are never shown as fact.
| Claim | Value | Status | Source |
|---|---|---|---|
facts.num_hidden_layers | 45 | • reported | hf_config · $.text_config.num_hidden_layers |
facts.hidden_size | 4096 | • reported | hf_config · $.text_config.hidden_size |
facts.num_attention_heads | 64 | • reported | hf_config · $.text_config.num_attention_heads |
facts.num_key_value_heads | — | ? unknown | |
facts.vocab_size | 154880 | • reported | hf_config · $.text_config.vocab_size |
facts.context_tokens | 1048576 | • reported | hf_config · $.text_config.max_position_embeddings |
facts.total_params | 320000000000 | • reported | model_card · parameter count |
facts.active_params | 18000000000 | = derived | derived · routed-active experts + shared + attention params |
topology.attention_groups[0] | "34 KDA layers (K,K,K,D schedule)" | • reported | hf_config · $.text_config.linear_attn_config.kda_layers |
topology.attention_groups[1] | "11 MLA/DSA layers" | • reported | hf_config · $.text_config.linear_attn_config.full_attn_layers |
topology.attention.kda_heads | 64 | • reported | hf_config · $.text_config.linear_attn_config.num_heads |
topology.attention.kda_head_dim | 128 | • reported | hf_config · $.text_config.linear_attn_config.head_dim |
topology.attention.kda_short_conv_kernel | 4 | • reported | hf_config · $.text_config.linear_attn_config.short_conv_kernel_size |
topology.attention.dsa_indexer_heads | 32 | • reported | hf_config · $.text_config.index_n_heads |
topology.attention.dsa_indexer_head_dim | 128 | • reported | hf_config · $.text_config.index_head_dim |
topology.attention.dsa_topk | 2048 | • reported | hf_config · $.text_config.index_topk |
topology.ffn_groups[0] | "first 3 blocks dense SwiGLU" | • reported | hf_config · $.text_config.first_k_dense_replace |
topology.ffn_groups[1] | "42 sparse MoE blocks" | • reported | hf_config · $.text_config.mlp_layer_types |
topology.experts.routed_total | 288 | • reported | hf_config · $.text_config.n_routed_experts |
topology.experts.active_routed | 8 | • reported | hf_config · $.text_config.num_experts_per_tok |
topology.experts.shared | 1 | • reported | hf_config · $.text_config.n_shared_experts |
topology.residual.streams | 4 | • reported | hf_config · $.text_config.hc_mult |
topology.residual.scheme | "mhc" | • reported | hf_config · $.text_config.mhc |
topology.mtp.predict_layers | 1 | • reported | hf_config · $.text_config.num_nextn_predict_layers |
Compare
Compare this model with any other catalog model (see catalog).