GLM · 2026-01 · Sparse MoE
GLM-4.7-Flash
GLM-4.7-Flash is a multi-head latent attention (MLA) transformer released by GLM in 2026-01, with 47 layers, hidden size 2048 and a context window of 202,752 tokens.
Layer stack
Attention ×47 · Layers 47
Key facts
| Family | GLM |
|---|---|
| Released | 2026-01 |
| Params | 29.4 B |
| Active | 3.2 B |
| Context | 202,752 tokens |
| Attention | MLA · kv_lora_rank 512 |
| Layers | 47 |
| Hidden | 2,048 |
| Heads | 20 |
| Vocab | 154,880 |
| Position | RoPE |
| Norm | RMSNorm |
| Activation | silu |
| Dtype | bf16 |
| Architecture class | Glm4MoeLiteForCausalLM |
Architecture overview
Attention
multi-head latent attention (MLA). kv_lora_rank=512. RoPE θ=1,000,000.
Feed-forward / MoE
It is a sparse mixture-of-experts with 64 routed experts and top-4 routing; about 3.2B parameters are active per token.
The configuration ships 1 extra multi-token-prediction (MTP) layers used to accelerate decoding.
Field-level comparison
Compared against the previous model of the same lab; where none exists, against the structurally closest model. The ratio column is this model divided by the comparison model.
Full-field comparison vs predecessor GLM-4.6
| Model | GLM-4.7-Flash | GLM-4.6 | Ratio |
|---|---|---|---|
| model_type | glm4_moe_lite | glm4_moe | ≠ |
| architectures | Glm4MoeLiteForCausalLM | Glm4MoeForCausalLM | ≠ |
| hidden_size | 2048 | 5120 | ×0.400 |
| num_hidden_layers | 47 | 92 | ×0.511 |
| num_attention_heads | 20 | 96 | ×0.208 |
| num_key_value_heads | 20 | 8 | ×2.50 |
| head_dim | — | 128 | |
| intermediate_size | 10240 | 12288 | ×0.833 |
| hidden_act | silu | silu | ＝ |
| n_routed_experts | 64 | 160 | ×0.400 |
| num_experts_per_tok | 4 | 8 | ×0.500 |
| n_shared_experts | 1 | 1 | ≈1 |
| moe_intermediate_size | 1536 | 1536 | ≈1 |
| kv_lora_rank | 512 | — | |
| qk_rope_head_dim | 64 | — | |
| q_lora_rank | 768 | — | |
| rope_theta | 1000000 | 1000000 | ≈1 |
| max_position_embeddings | 202752 | 202752 | ≈1 |
| rms_norm_eps | 0.00001 | 0.00001 | ≈1 |
| vocab_size | 154880 | 151552 | ×1.02 |
| torch_dtype | — | bfloat16 | |
| num_nextn_predict_layers | 1 | 1 | ≈1 |
| attention_bias | false | true | ≠ |
| attention_dropout | 0 | 0 | ＝ |
| pad_token_id | 154820 | 151329 | ×1.02 |
| eos_token_id | 154820×1 + 154827×1 + 154829×1 | 151329×1 + 151336×1 + 151338×1 | ≠ |
Most similar architectures
- DeepSeek-V2 Lite (DeepSeek, 2024-05) 0.998
- Kimi Linear 48B A3B (Kimi, 2025-10) 0.995
- Moonlight 16B (Kimi, 2025-02) 0.994
- DeepSeek-V2 (DeepSeek, 2024-05) 0.992
- DeepSeek-V3 (DeepSeek, 2024-12) 0.922
- DeepSeek-V3 Base (DeepSeek, 2024-12) 0.922
Similarity values range from 0 (no shared categorical features) to 1 (identical profiles).
Raw config fields
36 fields
| architectures | Glm4MoeLiteForCausalLM |
|---|---|
| attention_bias | false |
| attention_dropout | 0 |
| pad_token_id | 154820 |
| eos_token_id | 154820×1 + 154827×1 + 154829×1 |
| hidden_act | silu |
| hidden_size | 2048 |
| intermediate_size | 10240 |
| max_position_embeddings | 202752 |
| model_type | glm4_moe_lite |
| moe_intermediate_size | 1536 |
| topk_method | noaux_tc |
| norm_topk_prob | true |
| num_attention_heads | 20 |
| n_group | 1 |
| topk_group | 1 |
| n_routed_experts | 64 |
| n_shared_experts | 1 |
| routed_scaling_factor | 1.8 |
| num_experts_per_tok | 4 |
| first_k_dense_replace | 1 |
| num_hidden_layers | 47 |
| num_key_value_heads | 20 |
| num_nextn_predict_layers | 1 |
| partial_rotary_factor | 1 |
| rms_norm_eps | 0.00001 |
| rope_theta | 1000000 |
| tie_word_embeddings | false |
| dtype | bfloat16 |
| transformers_version | 5.0.0rc0 |
| q_lora_rank | 768 |
| kv_lora_rank | 512 |
| qk_nope_head_dim | 192 |
| qk_rope_head_dim | 64 |
| v_head_dim | 256 |
| vocab_size | 154880 |