RedHatAI/Qwen3-4B-Thinking-2507-quantized.w8a8

Qwen3-4B-Thinking-2507.w8a8

Model Overview

• Model Architecture: Qwen3ForCausalLM
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT8
• 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 the weights of Qwen/Qwen3-4B-Thinking-2507 to INT8 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. A combination of the SmoothQuant and GPTQ algorithms is applied for quantization, as implemented in the llm-compressor library.

Deployment

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/Qwen3-4B-Thinking-2507.w8a8"
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 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.

from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
from llmcompressor.transformers import oneshot
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model
model_stub = "Qwen/Qwen3-4B-Thinking"
model_name = model_stub.split("/")[-1]

num_samples = 1024
max_seq_len = 8192

model = AutoModelForCausalLM.from_pretrained(model_stub)

tokenizer = AutoTokenizer.from_pretrained(model_stub)

def preprocess_fn(example):
  return {"text": tokenizer.apply_chat_template(example["messages"], add_generation_prompt=False, tokenize=False)}

ds = load_dataset("neuralmagic/LLM_compression_calibration", split="train")
ds = ds.map(preprocess_fn)

# Configure the quantization algorithm and scheme
recipe = [
  SmoothQuantModifier(
    smoothing_strength=0.9,
    mappings=[
        [["re:.*q_proj", "re:.*k_proj", "re:.*v_proj"], "re:.*input_layernorm"],
    ],
  ),
  GPTQModifier(
        ignore: ["lm_head"]
        config_groups={"group_0": {"targets": ["Linear"], "weights": { "num_bits": 4, "type": int, "strategy": "group", "group_size": 128, "symmetric": true, "dynamic": false, "observer": "mse" } } },
        dampening_frac=0.1,
    )
]


# Apply quantization
oneshot(
    model=model,
    dataset=ds, 
    recipe=recipe,
    max_seq_length=max_seq_len,
    num_calibration_samples=num_samples,
)

# Save to disk in compressed-tensors format
save_path = model_name + "-quantized.w8a8"
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 ifeval, mmlu_pro and gsm8k_platinum using lm-evaluation-harness, on reasoning tasks using lighteval. vLLM was used for all evaluations.

Evaluation details

Deploy using vllm to create an OpenAI-compatible API endpoint:

• vLLM:

  vllm serve RedHatAI/Qwen3-4B-Thinking-2507.w8a8 --max-model-len 262144 --reasoning-parser deepseek_r1
  

  lm-evaluation-harness

  lm_eval --model local-chat-completions \
    --tasks mmlu_pro_chat \
    --model_args "model=RedHatAI/Qwen3-4B-Thinking-2507.w8a8,max_length=262144,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=64,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=1200" \
    --num_fewshot 0 \
    --apply_chat_template \
    --gen_kwargs "do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000
  

  lm_eval --model local-chat-completions \
    --tasks ifeval \
    --model_args "model=RedHatAI/Qwen3-4B-Thinking-2507.w8a8,max_length=262144,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=64,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=1200" \
    --num_fewshot 0 \
    --apply_chat_template \
    --gen_kwargs "do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000
  

  lm_eval --model local-chat-completions \
    --tasks gsm8k_platinum_cot_llama \
    --model_args "model=RedHatAI/Qwen3-4B-Thinking-2507.w8a8,max_length=262144,base_url=http://0.0.0.0:8000/v1/chat/completions,num_concurrent=64,max_retries=3,tokenized_requests=False,tokenizer_backend=None,timeout=1200" \
    --num_fewshot 0 \
    --apply_chat_template \
    --gen_kwargs "do_sample=True,temperature=0.6,top_p=0.95,top_k=20,min_p=0.0,max_gen_toks=64000
  

  lighteval

  lighteval_model_arguments.yaml

  model_parameters:
    model_name: RedHatAI/Qwen3-4B-Thinking-2507.w8a8
    dtype: auto
    gpu_memory_utilization: 0.9
    max_model_length: 40960
    generation_parameters:
      temperature: 0.6
      top_k: 20
      min_p: 0.0
      top_p: 0.95
      max_new_tokens: 32000
  

  lighteval endpoint litellm lighteval_model_arguments.yaml  \
    "aime25|0,math_500|0,gpqa:diamond|0"
  

Accuracy

Benchmark	Qwen3-4B Thinking	Qwen3-4B Thinking.w8a8 (this model)	Recovery (%)
GSM8k Platinum (0-shot)	95.26	95.78	100.55
MMLU-Pro (0-shot)	73.63	72.71	98.74
IfEval (0-shot)	90.17	89.81	99.60
AIME 2025	73.75	71.67	97.18
GPQA diamond	68.94	66.67	96.70
Math 500	89.45	89.28	99.80