Model Details
- Base Model: zai-org/GLM-5.2
- Architecture: glm_moe_dsa (GLM MoE with DeepSeek Sparse Attention)
- Total Parameters: 0.85B
- Activated Parameters: ~0.77B
Configuration Changes
The following parameters were reduced from the original model:
Table with columns: Parameter, Original, Tiny| Parameter | Original | Tiny |
|---|
| num_hidden_layers | 78 | 6 |
| hidden_size | 6144 | 2048 |
| intermediate_size | 12288 | 4096 |
| num_attention_heads | 64 | 16 |
| num_key_value_heads | 64 | 16 |
| n_routed_experts | 256 | 8 |
| num_experts_per_tok | 8 | 2 |
| moe_intermediate_size | 2048 | 512 |
| kv_lora_rank | 512 | 128 |
| q_lora_rank | 2048 | 512 |
| v_head_dim | 256 | 128 |
| index_n_heads | 32 | 8 |
| index_head_dim | 128 | 64 |
| first_k_dense_replace | 3 | 2 |
Checkpoint Structure
Single safetensors file containing 194 tensors in float32. Layers 0-1 have dense MLP, layers 2-5 have MoE MLP. Layers 0-2 have full DSA indexer weights, layers 3-5 use shared indexer.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("inference-optimization/GLM-5.2-0.8B-A0.8B", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("inference-optimization/GLM-5.2-0.8B-A0.8B")
input_ids = tokenizer("According to all known laws", return_tensors="pt").input_ids.to(model.device)
output = model.generate(input_ids, max_new_tokens=20)
print(tokenizer.decode(output[0]))
Creation Process
- Inspected original GLM-5.2 config (78 layers, 256 experts, hidden_size=6144)
- Reduced all dimensions to target ~1B parameters while preserving architecture
- Created model with float32 dtype for training stability
- Fine-tuned on copypasta dataset to perplexity ~1.0
- Validated checkpoint structure matches original model naming conventions
- Validated model loads, inferences, and generates correctly
Validation Output
Success: 1.0000379085540771 <= 10.0
Generating sample text:
According to all known laws of aviation, there is no way a bee should be able to fly.
Notes
- The model uses float32 dtype (original uses bfloat16) to ensure proper initialization and training of the tiny model
- Architecture preserves both dense and sparse MLP layer types, MLA attention with compressed Q/KV, and DSA indexer with full/shared patterns
- The model has been fine-tuned on a toy dataset and is intended for testing purposes only