inference-optimization / Qwen3-1.6B-A0.9B

Qwen3-1.6B-A0.9B

This is a tiny version of Qwen/Qwen3-30B-A3B created for testing and development.

Model Details

• Base Model : Qwen/Qwen3-30B-A3B
• Architecture : qwen3_moe (Mixture of Experts)
• Total Parameters : 1.57B
• Activated Parameters : ~0.9B (50% MoE activation)

Configuration Changes

The following parameters were reduced from the original model:

Checkpoint Structure

The checkpoint is stored as a single model.safetensors file with individual expert weights matching the original Qwen3 structure. Each layer has 16 experts with separate gate_proj , up_proj , and down_proj weights per expert.

Validation

The model was fine-tuned on a toy copypasta dataset and achieves:

• Perplexity : 1.0 (on validation text)
• Generation : Successfully generates coherent continuations

Example generation:

Input: "According to all known laws"
Output: "According to all known laws of aviation, there is no way a bee should be able to fly."

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("inference-optimization/Qwen3-1.6B-A0.9B", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("inference-optimization/Qwen3-1.6B-A0.9B")

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

This model was created using the llm-compressor create-tiny-model Claude skill:

1. Configuration : Created with 10 layers and 16 experts (8 activated per token) to achieve ~1.6B total parameters with 50% MoE activation
2. Initialization : Randomly initialized weights using transformers init_weights()
3. Fine-tuning : Trained on famous internet copypastas until perplexity < 3.0
4. Checkpoint Conversion : Converted from batched expert format to individual expert format to match original Qwen3 checkpoint structure
5. Validation : Confirmed perplexity ~1.0 and successful text generation

Notes

• This model maintains the same MoE architecture as the original with Grouped Query Attention (GQA)
• The checkpoint format exactly matches the original Qwen3-30B-A3B structure for compatibility
• This model is intended for testing and development only, not production use