Qwen3-VL-30B-A3B-Thinking.w4a16

Model Overview

• Model Optimizations:
  • Weight quantization: INT4

Model Optimizations

This model was obtained by quantizing the weights of Qwen/Qwen3-VL-30B-A3B-Thinking to INT4 data type. This optimization reduces the number of bits per parameter from 16 to 4, reducing the disk size and GPU memory requirements by approximately 75%.

Only the weights of the linear operators within transformers blocks are quantized. Weights are quantized using a symmetric per-group scheme, with group size 128. The AutoRound algorithm is applied for quantization, as implemented as AutoRoundModifier in the llm-compressor library.

Deployment

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

vllm serve EmbeddedLLM/Qwen3-VL-30B-A3B-Thinking.w4a16 --reasoning_parser deepseek_r1

Creation

Creation details

This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.

from auto_round.calib_dataset import get_dataset
from transformers import AutoTokenizer, Qwen3VLMoeForConditionalGeneration, AutoProcessor

from llmcompressor import oneshot
from llmcompressor.modifiers.autoround import AutoRoundModifier
from llmcompressor.utils import dispatch_for_generation

# Select model and load it.
model_id = "Qwen/Qwen3-VL-30B-A3B-Thinking"
model = Qwen3VLMoeForConditionalGeneration.from_pretrained(model_id, dtype="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)


# Select calibration dataset.
NUM_CALIBRATION_SAMPLES = 128
MAX_SEQUENCE_LENGTH = 2048
# Get aligned calibration dataset.
ds = get_dataset(
    tokenizer=tokenizer,
    seqlen=MAX_SEQUENCE_LENGTH,
    nsamples=NUM_CALIBRATION_SAMPLES,
)


# Configure the quantization algorithm to run.
#   * quantize the weights to 4 bit with AutoRound with a group size 128
recipe = AutoRoundModifier(
    targets="Linear",
    scheme="W4A16",
    ignore=[
        "re:.*lm_head",
        "re:visual.*",
        "re:model.visual.*",
        "re:.*mlp.gate$",
    ],
    iters=200
)

oneshot(
  model=model,
  dataset=ds,
  recipe=recipe,
  max_seq_length=MAX_SEQUENCE_LENGTH,
  num_calibration_samples=NUM_CALIBRATION_SAMPLES,
  # disable shuffling to get slightly better mmlu score
  shuffle_calibration_samples=False,
)

# Save to disk compressed.
SAVE_DIR = model_id.rstrip("/").split("/")[-1] + ".w4a16"
model.save_pretrained(SAVE_DIR, save_compressed=True)
tokenizer.save_pretrained(SAVE_DIR)
processor.save_pretrained(SAVE_DIR)

Evaluation

After started the vllm server as shown above, the model was evaluated on with the following commands.

Disclaimer: Results may differ from official benchmarks due to evaluation setup variations.

Evaluation details

lm-evaluation-harness

lm_eval
  --model local-chat-completions \
  --model_args model="EmbeddedLLM/Qwen3-VL-30B-A3B-Thinking.w4a16",base_url=http://127.0.0.1:8000/v1/chat/completions,num_concurrent=64,timeout=300 \
  --gen_kwargs max_gen_toks=8196 \
  --tasks mmlu_pro \
  --apply_chat_template

lmms-eval

lmms-eval \
  --model async_openai \
  --model_args model_version=EmbeddedLLM/Qwen3-VL-30B-A3B-Thinking.w4a16,base_url=http://127.0.0.1:8000/v1,is_qwen3_vl=True,api_key=DUMMY,num_cpus=8,timeout=300 \
  --gen_kwargs max_new_tokens=32768 \
  --tasks mmmu_val

Accuracy

Category	Benchmark	Qwen3-VL-30B-A3B-Thinking	Qwen3-VL-30B-A3B-Thinking.w4a16 (this model)	Recovery
Text	MMLU_pro	66.96	65.86	98.4%
Vision	MMMU_val	63	63	100.0%