GLM · 2026-08 · Sparsames MoE · Multimodal
GLM-5.3-Flash
GLM-5.3-Flash ist ein Multi-Head-Latent-Attention (MLA)-Transformer, veröffentlicht von GLM im 2026-08, mit 45 Schichten, Hidden-Größe 4096 und einem Kontextfenster von 1,048,576 Tokens.
Schichtstapel
Kernfakten
| Labor | GLM |
|---|---|
| Veröffentlicht | 2026-08 |
| Parameter | 328.9 Mrd. |
| Aktiv | 13 Mrd. |
| Kontext | 1,048,576 tokens |
| Attention | MLA · kv_lora_rank 512 |
| Schichten | 45 |
| Hidden-Dim. | 4,096 |
| Attention-Köpfe | 64 |
| Vokabular | 154,880 |
| Positionscodierung | learned/absolute |
| Normierung | RMSNorm |
| Aktivierung | silu |
| Präzision | fp8 |
| Architekturklasse | Glm5NextForConditionalGeneration |
Architekturüberblick
Attention
Multi-Head-Latent-Attention (MLA). kv_lora_rank=512.
FFN / MoE
Es handelt sich um ein sparsames Mixture-of-Experts mit 288 gerouteten Experten und Top-8-Routing; etwa 13 Mrd. Parameter sind pro Token aktiv.
Die Konfiguration enthält 1 zusätzliche Multi-Token-Prediction-Schichten (MTP) zur Beschleunigung der Dekodierung.
Feldweiser Vergleich
Verglichen mit dem vorherigen Modell desselben Labors; fehlt dieses, mit dem strukturell nächsten Modell. Die Spalte Ratio = dieses Modell ÷ Vergleichsmodell.
Vollfeldvergleich mit dem Vorgänger GLM-5.3
| Modell | GLM-5.3-Flash | GLM-5.3 | Ratio |
|---|---|---|---|
| model_type | glm5_next | glm_moe_dsa | ≠ |
| architectures | Glm5NextForConditionalGeneration | GlmMoeDsaForCausalLM | ≠ |
| hidden_size | 4096 | 6144 | ×0.667 |
| num_hidden_layers | 45 | 78 | ×0.577 |
| num_attention_heads | 64 | 64 | ≈1 |
| num_key_value_heads | 64 | 64 | ≈1 |
| head_dim | 0 | 192 | ×0.000 |
| intermediate_size | 12288 | 12288 | ≈1 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | 288 | 256 | ×1.13 |
| num_experts_per_tok | 8 | 8 | ≈1 |
| n_shared_experts | 1 | 1 | ≈1 |
| moe_intermediate_size | 2048 | 2048 | ≈1 |
| kv_lora_rank | 512 | 512 | ≈1 |
| qk_rope_head_dim | 0 | 64 | ×0.000 |
| q_lora_rank | 1536 | 2048 | ×0.750 |
| max_position_embeddings | 1048576 | 1048576 | ≈1 |
| rms_norm_eps | 0.00001 | 0.00001 | ≈1 |
| vocab_size | 154880 | 154880 | ≈1 |
| num_nextn_predict_layers | 1 | 1 | ≈1 |
| image_token_id | 154854 | — | |
| video_token_id | 154855 | — | |
| image_start_token_id | 154830 | — | |
| image_end_token_id | 154831 | — | |
| video_start_token_id | 154832 | — | |
| video_end_token_id | 154833 | — |
Strukturell ähnlichste Modelle
- GLM-5.2 (GLM, 2026-06) 0.994
- GLM-5.3 (GLM, 2026-08) 0.994
- Hunyuan 4 Preview (Hunyuan, 2026-08) 0.994
- GLM-5 (GLM, 2026-02) 0.992
- GLM-5.1 (GLM, 2026-04) 0.992
- Kimi K2 0905 (Kimi, 2025-09) 0.935
Ähnlichkeitswerte reichen von 0 (keine gemeinsamen kategorialen Merkmale) bis 1 (identische Profile).
Rohe Config-Felder
65 Felder
| architectures | Glm5NextForConditionalGeneration |
|---|---|
| image_token_id | 154854 |
| video_token_id | 154855 |
| image_start_token_id | 154830 |
| image_end_token_id | 154831 |
| video_start_token_id | 154832 |
| video_end_token_id | 154833 |
| tie_word_embeddings | false |
| model_type | glm5_next |
| transformers_version | 5.16.0 |
| attention_bias | false |
| attention_dropout | 0 |
| dtype | bfloat16 |
| eos_token_id | 154820×1 + 154827×1 + 154829×1 |
| first_k_dense_replace | 3 |
| hc_eps | 0.000001 |
| hc_mult | 4 |
| hc_sinkhorn_iters | 20 |
| head_dim | 0 |
| hidden_act | silu |
| hidden_size | 4096 |
| index_head_dim | 128 |
| index_kpool | 4 |
| index_kpool_always_select_tail | true |
| index_kpool_compress | true |
| index_n_heads | 32 |
| index_topk | 2048 |
| index_share_for_mtp_iteration | true |
| indexer_rope_interleave | true |
| indexer_types | full×45 |
| initializer_range | 0.02 |
| intermediate_size | 12288 |
| kv_lora_rank | 512 |
| layer_types | linear_attention×34 + deepseek_sparse_attention×11 |
| max_position_embeddings | 1048576 |
| mhc | true |
| mla_use_nope | true |
| mlp_layer_types | dense×3 + sparse×42 |
| moe_intermediate_size | 2048 |
| moe_router_dtype | float32 |
| n_group | 1 |
| n_routed_experts | 288 |
| n_shared_experts | 1 |
| norm_topk_prob | true |
| num_attention_heads | 64 |
| num_experts_per_tok | 8 |
| num_hidden_layers | 45 |
| num_key_value_heads | 64 |
| num_nextn_predict_layers | 1 |
| output_router_logits | false |
| pad_token_id | 154820 |
| q_lora_rank | 1536 |
| qk_head_dim | 256 |
| qk_nope_head_dim | 256 |
| qk_rope_head_dim | 0 |
| rms_norm_eps | 0.00001 |
| routed_scaling_factor | 2.5 |
| router_aux_loss_coef | 0.001 |
| scoring_func | sigmoid |
| swiglu_limit | 10 |
| topk_group | 1 |
| topk_method | noaux_tc |
| use_cache | true |
| v_head_dim | 256 |
| vocab_size | 154880 |