Model Architecture: Qwen3MoeForCausalLM
- Input: Text
- Output: Text
Model Optimizations:
- Activation quantization: FP8
- Weight quantization: FP8
Intended Use Cases:
- Reasoning.
- Function calling.
- Subject matter experts via fine-tuning.
- Multilingual instruction following.
- Translation.
Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws).
Release Date: 05/05/2025
Version: 1.0
Model Developers: RedHat (Neural Magic)

Model Optimizations

This model was obtained by quantizing activations and weights of Qwen3-235B-A22B to FP8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size requirements by approximately 50%.

Only weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized with a symmetric static per-channel scheme, whereas activations are quantized with a symmetric dynamic per-token scheme. The llm-compressor library is used for quantization.

Deployment

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

python

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "RedHatAI/Qwen3-235B-A22B-FP8-dynamic"
number_gpus = 4
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, top_k=20, min_p=0, max_tokens=256)

messages = [
    {"role": "user", "content": prompt}
]

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [{"role": "user", "content": "Give me a short introduction to large language model."}]

prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

llm = LLM(model=model_id, tensor_parallel_size=number_gpus)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

vLLM aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

Creation details

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

python

from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model
model_stub = "Qwen/Qwen3-235B-A22B"
model_name = model_stub.split("/")[-1]

model = AutoModelForCausalLM.from_pretrained(model_stub)

tokenizer = AutoTokenizer.from_pretrained(model_stub)

# Configure the quantization algorithm and scheme
recipe = QuantizationModifier(
    ignore=["lm_head"],
    targets="Linear",
    scheme="FP8_dynamic",
)

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

# Save to disk in compressed-tensors format
save_path = model_name + "-FP8-dynamic"
model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)
print(f"Model and tokenizer saved to: {save_path}")

Evaluation

The model was evaluated on the OpenLLM leaderboard tasks (version 1), using lm-evaluation-harness and vLLM.

Evaluation details

lm_eval \
  --model vllm \
  --model_args pretrained="RedHatAI/Qwen3-235B-A22B-FP8-dynamic",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=4 \
  --tasks openllm \
  --apply_chat_template\
  --fewshot_as_multiturn \
  --batch_size auto

Accuracy

Category	Benchmark	Qwen3-235B-A22B	Qwen3-235B-A22B-FP8-dynamic (this model)	Recovery
OpenLLM v1	MMLU (5-shot)	84.77	84.61	99.8%
ARC Challenge (25-shot)	71.84	70.90	98.7%
GSM-8K (5-shot, strict-match)	74.22	74.98	101.0%
Hellaswag (10-shot)	76.56	76.10	99.4%
Winogrande (5-shot)	73.95	75.06	101.5%
TruthfulQA (0-shot, mc2)	61.18	60.93	99.6%
Average	73.75	73.76	100.0%