temporary-3-layer-Qwen3-VL-1.0B-A0.4B-Instruct

Qwen3-VL-1.0B-A0.4B-Instruct

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

Model Details

• Base Model: Qwen/Qwen3-VL-30B-A3B-Instruct
• Architecture: qwen3_vl_moe
• Total Parameters: 1.434B
• Activated Parameters: ~0.4B (8 of 128 experts active per token)

Configuration Changes

The following parameters were reduced from the original model:

Checkpoint Structure

The model is saved as a single model.safetensors file (89 tensors). The checkpoint key structure matches the original Qwen/Qwen3-VL-30B-A3B-Instruct exactly.

Usage

from transformers import Qwen3VLMoeForConditionalGeneration, AutoTokenizer

model = Qwen3VLMoeForConditionalGeneration.from_pretrained(
    "inference-optimization/Qwen3-VL-1.0B-A0.4B-Instruct",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("inference-optimization/Qwen3-VL-1.0B-A0.4B-Instruct")

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]))

Note: AutoModelForCausalLM does not support this model type. Use Qwen3VLMoeForConditionalGeneration or AutoModelForImageTextToText directly.

Creation Process

This model was created using the llm-compressor create-tiny-model claude skill.

1. Config fetched from Qwen/Qwen3-VL-30B-A3B-Instruct without downloading weights (skip_weights_download)
2. Text depth reduced from 48 → 1 layers; vision depth reduced from 27 → 4 blocks
3. All parameters randomly initialized with initializer_range=0.02 (matching the base model)
4. MoE expert parameters (gate_up_proj, down_proj, gate.weight) explicitly initialized since they are bare nn.Parameter tensors, not nn.Linear modules
5. Fine-tuned on a toy text dataset (internet copypastas) until perplexity ≤ 3.0

Validation

Success: perplexity=1.0001 <= 10.0
Generated: According to all known laws of aviation, there is no way a bee should be able to fly. Its wings are too small to get its fat little

Notes

• This is a text-only fine-tuned model. The vision encoder weights are randomly initialized and not fine-tuned; the model is intended for text-based testing only.
• The MoE routing architecture (128 experts, 8 active per token) is preserved in full to match the original model's structural properties.
• Fine-tuning converged in ~160 steps with learning rate 1e-4.