đ Qwen3-30B-A3B-FP8-dynamic
This is a text generation model based on the Qwen3 architecture, optimized through FP8 quantization.
đ Quick Start
This model can be deployed efficiently using the vLLM backend, as shown in the example below.
from vllm import LLM, SamplingParams
from transformers import AutoTokenizer
model_id = "RedHatAI/Qwen3-30B-A3B-FP8-dynamic"
number_gpus = 1
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 also supports OpenAI-compatible serving. See the documentation for more details.
⨠Features
- Model Architecture: Qwen3MoeForCausalLM, taking text as input and outputting 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 Qwen3-30B-A3B 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.
đ Documentation
Creation
Creation details
This model was created with [llm-compressor](https://github.com/vllm-project/llm-compressor) by running the code snippet below.
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot
from transformers import AutoModelForCausalLM, AutoTokenizer
model_stub = "Qwen/Qwen3-30B-A3B"
model_name = model_stub.split("/")[-1]
model = AutoModelForCausalLM.from_pretrained(model_stub)
tokenizer = AutoTokenizer.from_pretrained(model_stub)
recipe = QuantizationModifier(
ignore=["lm_head"],
targets="Linear",
scheme="FP8_dynamic",
)
oneshot(
model=model,
recipe=recipe,
)
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/Qwen3-30B-A3B-FP8-dynamic",dtype=auto,gpu_memory_utilization=0.5,max_model_len=8192,enable_chunk_prefill=True,tensor_parallel_size=1 \
--tasks openllm \
--apply_chat_template\
--fewshot_as_multiturn \
--batch_size auto
Accuracy
Category |
Benchmark |
Qwen3-30B-A3B |
Qwen3-30B-A3B-FP8-dynamic (this model) |
Recovery |
OpenLLM v1 |
MMLU (5-shot) |
77.67 |
77.49 |
99.8% |
OpenLLM v1 |
ARC Challenge (25-shot) |
63.40 |
63.65 |
100.4% |
OpenLLM v1 |
GSM-8K (5-shot, strict-match) |
87.26 |
86.73 |
99.4% |
OpenLLM v1 |
Hellaswag (10-shot) |
54.33 |
54.33 |
100.0% |
OpenLLM v1 |
Winogrande (5-shot) |
66.77 |
66.30 |
99.3% |
OpenLLM v1 |
TruthfulQA (0-shot, mc2) |
56.27 |
56.88 |
101.1% |
OpenLLM v1 |
Average |
67.62 |
67.56 |
99.9% |
OpenLLM v2 |
MMLU-Pro (5-shot) |
47.45 |
48.40 |
102.0% |
OpenLLM v2 |
IFEval (0-shot) |
86.26 |
86.08 |
99.8% |
OpenLLM v2 |
BBH (3-shot) |
34.81 |
34.70 |
99.7% |
OpenLLM v2 |
Math-lvl-5 (4-shot) |
52.14 |
59.39 |
113.9% |
OpenLLM v2 |
GPQA (0-shot) |
0.31 |
0.90 |
--- |
OpenLLM v2 |
MuSR (0-shot) |
8.09 |
9.05 |
--- |
OpenLLM v2 |
Average |
38.18 |
39.75 |
104.1% |
Multilingual |
MGSM (0-shot) |
32.27 |
32.73 |
101.5% |
Reasoning (generation) |
AIME 2024 |
78.33 |
78.96 |
100.8% |
Reasoning (generation) |
AIME 2025 |
71.46 |
68.44 |
95.8% |
Reasoning (generation) |
GPQA diamond |
62.63 |
62.63 |
100.0% |
Reasoning (generation) |
Math-lvl-5 |
97.60 |
95.80 |
98.2% |
Reasoning (generation) |
LiveCodeBench |
60.66 |
60.89 |
100.4% |
đ License
This project is licensed under the Apache-2.0 license.