GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
notSnix
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
notSnix
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
| teacher | this model | |
|---|---|---|
| params (LM body) | 27.4B | 16.6B |
| memory (batch 96, 4k+4k) | 97.5 GiB | 77.9 GiB |
| lm loss | 2.265 | 3.026 |
| token acc top-1 | 0.688 | 0.588 |
| token acc top-5 | 0.877 | 0.780 |
| token acc top-10 | 0.907 | 0.822 |
FFN-only, v1 scope. Both mixer kinds (Gated DeltaNet linear_attn and full self_attn) are the
teacher's, unchanged. Vision tower carried through unchanged. MTP head dropped. Per-layer FFN
intermediate size chosen by MIP from {4352, 8704, 13056, 17408}:
markdown
block_0: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_1: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_2: mamba num_heads_48 head_dim_128 ffn intermediate_8704block_3: attention kv_heads_4 ffn intermediate_4352block_4: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_5: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_6: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_7: attention kv_heads_4 ffn no_opblock_8: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_9: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_10: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_11: attention kv_heads_4 ffn intermediate_4352block_12: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_13: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_14: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_15: attention kv_heads_4 ffn intermediate_4352block_16: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_17: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_18: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_19: attention kv_heads_4 ffn intermediate_17408block_20: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_21: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_22: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_23: attention kv_heads_4 ffn intermediate_17408block_24: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_25: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_26: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_27: attention kv_heads_4 ffn intermediate_17408block_28: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_29: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_30: mamba num_heads_48 head_dim_128 ffn intermediate_17408block_31: attention kv_heads_4 ffn intermediate_4352block_32: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_33: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_34: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_35: attention kv_heads_4 ffn intermediate_4352block_36: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_37: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_38: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_39: attention kv_heads_4 ffn intermediate_4352block_40: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_41: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_42: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_43: attention kv_heads_4 ffn intermediate_4352block_44: mamba num_heads_48 head_dim_128 ffn no_opblock_45: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_46: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_47: attention kv_heads_4 ffn intermediate_4352block_48: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_49: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_50: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_51: attention kv_heads_4 ffn intermediate_4352block_52: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_53: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_54: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_55: attention kv_heads_4 ffn intermediate_4352block_56: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_57: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_58: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_59: attention kv_heads_4 ffn intermediate_4352block_60: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_61: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_62: mamba num_heads_48 head_dim_128 ffn intermediate_4352block_63: attention kv_heads_4 ffn intermediate_4352
AnyModel subblock-safetensors (subblocks_safetensors/ + model.safetensors.index.json) with
block_configs in config.json. Load with NVIDIA Model-Optimizer's puzzletron tooling
(modelopt.torch.puzzletron) using the qwen3_5 descriptor. Requires transformers >= 5.8 and the
qwen3_5 descriptor patch (see the companion data repo notSnix/Qwen3.8-27B-Puzzletron-data).