Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK

Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK

Model Overview

• Model Architecture: Qwen3VLMoeForConditionalGeneration
  • Input: Text, Image
  • Output: Text
• Model Optimizations:
  • Weight quantization: FP8
  • Activation quantization: FP8
• Release Date:
• Version: 1.0
• Model Developers:: Red Hat

Quantized version of Qwen/Qwen3-VL-235B-A22B-Instruct.

Model Optimizations

This model was obtained by quantizing the weights and activations of Qwen/Qwen3-VL-235B-A22B-Instruct to FP8 data type. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%. Only the weights and activations of the linear operators within transformers blocks of the language model are quantized.

Deployment

Use with vLLM

1. Initialize vLLM server:

vllm serve nm-testing/Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK --tensor_parallel_size 8

1. Send requests to the server:

from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://<your-server-host>:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

model = "nm-testing/Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"},
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

outputs = client.chat.completions.create(
    model=model,
    messages=messages,
)

generated_text = outputs.choices[0].message.content
print(generated_text)

Creation

This model was quantized using the llm-compressor library as shown below.

<details> <summary>Creation details</summary>

from transformers import AutoProcessor, Qwen3VLMoeForConditionalGeneration

from llmcompressor import oneshot
from llmcompressor.modeling import replace_modules_for_calibration
from llmcompressor.modifiers.quantization import QuantizationModifier

MODEL_ID = "Qwen/Qwen3-VL-235B-A22B-Instruct"

# Load model.
model = Qwen3VLMoeForConditionalGeneration.from_pretrained(MODEL_ID, dtype="auto")
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = replace_modules_for_calibration(model)

# Configure the quantization algorithm and scheme.
# In this case, we:
#   * quantize the weights to fp8 with per-block quantization
#   * quantize the activations to fp8 with dynamic token activations
recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_BLOCK",
    ignore=[
        "re:.*lm_head",
        "re:visual.*",
        "re:model.visual.*",
        "re:.*mlp.gate$",
    ],
)

# Apply quantization.
oneshot(model=model, recipe=recipe)

# Save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.rstrip("/").split("/")[-1] + "-FP8-block"
model.save_pretrained(SAVE_DIR)
processor.save_pretrained(SAVE_DIR)

</details>

Evaluation

The model was evaluated on the OpenLLMv1 leaderboard task, using lm-evaluation-harness, on reasoning tasks using lighteval. vLLM was used for all evaluations.

<details> <summary>Evaluation details</summary>

Openllm V1

lm_eval \
  --model vllm \
  --model_args pretrained="nm-testing/Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK",dtype=auto,add_bos_token=True,max_model_len=16384,tensor_parallel_size=4,gpu_memory_utilization=0.9,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks openllm \
  --write_out \
  --batch_size auto \
  --output_path $output_path/openllm.json \
  --show_config

Openllm V2

lm_eval \
  --model vllm \
  --model_args pretrained="nm-testing/Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK",dtype=auto,add_bos_token=False,max_model_len=16384,tensor_parallel_size=4,gpu_memory_utilization=0.7,disable_log_stats=True,enable_chunked_prefill=True,trust_remote_code=True \
  --tasks leaderboard \
  --apply_chat_template \
  --fewshot_as_multiturn \
  --write_out \
  --batch_size auto \
  --output_path $output_path/leaderboard.json \
  --show_config

Coding Benchmarks

evalplus.evaluate --model "nm-testing/Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK" \
                  --dataset "humaneval" \
                  --backend vllm \
                  --tp 4 \
                  --greedy
evalplus.evaluate --model "nm-testing/Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK" \
                --dataset "mbpp" \
                --backend vllm \
                --tp 4 \
                --greedy

</details>

Accuracy

<table> <thead> <tr> <th>Category</th> <th>Metric</th> <th>Qwen/Qwen3-VL-235B-A22B-Instruct</th> <th>nm-testing/Qwen3-VL-235B-A22B-Instruct-FP8-BLOCK</th> <th>Recovery (%)</th> </tr> </thead> <tbody> <!-- OpenLLM Leaderboard V1 --> <tr> <td rowspan="7"><b>OpenLLM V1</b></td> <td>ARC-Challenge (Acc-Norm, 25-shot)</td> <td>76.19</td> <td>76.28</td> <td>100.11</td> </tr> <tr> <td>GSM8K (Strict-Match, 5-shot)</td> <td>41.24</td> <td>41.70</td> <td>101.10</td> </tr> <tr> <td>HellaSwag (Acc-Norm, 10-shot)</td> <td>87.89</td> <td>87.65</td> <td>99.73</td> </tr> <tr> <td>MMLU (Acc, 5-shot)</td> <td>87.15</td> <td>87.25</td> <td>100.11</td> </tr> <tr> <td>TruthfulQA (MC2, 0-shot)</td> <td>63.08</td> <td>63.24</td> <td>100.26</td> </tr> <tr> <td>Winogrande (Acc, 5-shot)</td> <td>82.00</td> <td>81.85</td> <td>99.81</td> </tr> <tr> <td><b>Average Score</b></td> <td><b>72.92</b></td> <td><b>73.00</b></td> <td><b>100.11</b></td> </tr> <!-- OpenLLM Leaderboard V2 --> <tr> <td rowspan="7"><b>OpenLLM V2</b></td> <td>IFEval (Inst Level Strict Acc, 0-shot)</td> <td>91.01</td> <td>90.29</td> <td>99.21</td> </tr> <tr> <td>BBH (Acc-Norm, 3-shot)</td> <td>73.72</td> <td>73.95</td> <td>100.31</td> </tr> <tr> <td>Math-Hard (Exact-Match, 4-shot)</td> <td>61.71</td> <td>20.69</td> <td>33.54</td> </tr> <tr> <td>GPQA (Acc-Norm, 0-shot)</td> <td>32.13</td> <td>32.89</td> <td>102.35</td> </tr> <tr> <td>MUSR (Acc-Norm, 0-shot)</td> <td>42.06</td> <td>41.80</td> <td>99.37</td> </tr> <tr> <td>MMLU-Pro (Acc, 5-shot)</td> <td>65.82</td> <td>65.65</td> <td>99.73</td> </tr> <tr> <td><b>Average Score</b></td> <td><b>61.07</b></td> <td><b>54.21</b></td> <td><b>88.77</b></td> </tr> </tbody> </table>