Qwen3-4B-Instruct-2507-heretic-W8A16

Overview

This is an 8-bit weight-only quantized version of Qwen3-4B-Instruct-2507-heretic using the W8A16 scheme, quantized with LLMCompressor.

Property	Value
Base Model	Qwen3-4B-Instruct-2507-heretic
Quantization	W8A16 (8-bit weights, 16-bit activations)
Quant Method	compressed-tensors (AWQ-compatible)
Model Size	~4.1 GB
Context Length	262,144 tokens

Quantization Details

• Framework: LLMCompressor
• Quantization Scheme: W8A16 (channel-wise symmetric int8)
• Format: pack-quantized (compressed-tensors)
• Target Layers: All Linear layers (except lm_head)

Performance

Perplexity (WikiText-2-raw-v1, test split)

Model	Perplexity	Degradation
Original FP16	17.4634	-
This Model (W8A16)	17.5556	+0.53%

Nearly lossless quantization with minimal perplexity degradation!

Usage

With vLLM

vllm serve groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16 \
  --quantization compressed-tensors \
  --trust-remote-code

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16"
)

messages = [{"role": "user", "content": "Hello, how are you?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Python API (vLLM)

from vllm import LLM, SamplingParams

llm = LLM(
    model="groxaxo/Qwen3-4B-Instruct-2507-heretic-W8A16",
    quantization="compressed-tensors",
    trust_remote_code=True
)

sampling_params = SamplingParams(temperature=0.7, top_p=0.9, max_tokens=256)
outputs = llm.generate(["Hello, how are you?"], sampling_params)
for output in outputs:
    print(output.outputs[0].text)

Model Architecture

• Architecture: Qwen3ForCausalLM
• Hidden Size: 2,560
• Intermediate Size: 9,728
• Attention Heads: 32
• KV Heads: 8 (GQA)
• Layers: 36
• Vocab Size: 151,936
• RoPE Theta: 5,000,000

Acknowledgements

• Qwen Team for the original Qwen3 model
• heretic-org for the fine-tuned variant
• vLLM Team for LLMCompressor and vLLM

License

Apache 2.0 (inherited from base model)