Qwen3-4B-Computer-Science-AWQ

Overview

Qwen3-4B-Computer-Science-AWQ is the AWQ-calibrated quantized release of Qwen3-4B-Computer-Science.

Weights are stored using the Compressed-Tensors format with 4-bit asymmetric group-wise quantization (W4A16). The checkpoint was produced using Activation-aware Weight Quantization (AWQ) calibration and validated by successful quantization, checksum verification, and CPU inference.

---

Model Information

Property	Value
Base Model	Qwen/Qwen3-4B
Model Family	Qwen3-4B-Computer-Science
Quantization	AWQ
Storage Format	Compressed-Tensors
Weight Precision	INT4
Activation Precision	FP16 / BF16
Quantization Scheme	W4A16
Group Size	128
Weight Quantization	Asymmetric
lm_head	Excluded from Quantization
Language	English
License	Apache-2.0

---

Training Data

The base model was instruction tuned using permissively licensed datasets.

Dataset	Configuration	License
HuggingFaceTB/smoltalk	smol-magpie-ultra	Apache-2.0
agentica-org/DeepCoder-Preview-Dataset	primeintellect	MIT

Dataset Size

Split	Samples
Training	60,989
Evaluation	512

---

Intended Use

This model is intended for:

• Software engineering
• Programming
• Code generation
• Debugging
• Code review
• Algorithm implementation
• Computer science education
• General technical reasoning

---

Quantization

This release was generated using Activation-aware Weight Quantization (AWQ).

The resulting checkpoint stores weights using packed 4-bit group-wise asymmetric quantization.

Parameter	Value
Weight Format	Packed INT4
Group Size	128
Symmetric	No
Observer	memoryless_minmax
Compression Format	Compressed-Tensors

---

Runtime Compatibility

This checkpoint uses the Compressed-Tensors format.

It is intended for runtimes that support Compressed-Tensors models.

Validation performed for this release:

• Successful AWQ calibration
• Successful model serialization
• CPU inference
• SHA256 verification of release artifacts

Loading this checkpoint with standard Transformers may decompress weights during execution depending on the runtime and available hardware.

---

Usage

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "Irfanuruchi/Qwen3-4B-Computer-Science-AWQ",
    device_map="auto",
    dtype="auto",
)

tokenizer = AutoTokenizer.from_pretrained(
    "Irfanuruchi/Qwen3-4B-Computer-Science-AWQ"
)

---

Release Artifacts

model.safetensors
config.json
generation_config.json
recipe.yaml
tokenizer.json
tokenizer_config.json
chat_template.jinja
SHA256SUMS
LICENSE
README.md

---

Integrity Verification

Every release artifact includes a SHA256 checksum.

Verify downloaded files:

sha256sum -c SHA256SUMS

---

Validation

The published checkpoint was verified before release.

Completed validation:

• AWQ calibration completed successfully
• Quantized checkpoint generated successfully
• CPU inference completed successfully
• SHA256 checksums verified

---

Limitations

• Quantization may affect output quality compared to the BF16 checkpoint.
• Runtime support depends on the inference engine.
• GPU memory requirements depend on whether the runtime executes directly on compressed weights or decompresses them during inference.

---

License

Base model:

• Apache-2.0

Training datasets:

• Apache-2.0
• MIT

This repository is distributed under the Apache-2.0 License.

---

Acknowledgements

• Alibaba Qwen Team
• Hugging Face
• vLLM Project
• LLM Compressor Project
• SmolTalk Contributors
• DeepCoder Contributors

---

Citation

@software{uruci2026qwen3csawq,
  title={Qwen3-4B-Computer-Science-AWQ},
  author={Irfan Uruçi},
  year={2026},
  publisher={Hugging Face}
}