Qwen3.5-9B-EOQ-v3

🏆 EOQ v3 -- Qwen3.5-9B (PolarQuant + AWQ)

Near-lossless quantization: PPL 6.43 -- only +0.06 from FP16 6.37. The best quality result in the PolarQuant family.

EOQ v3 combines PolarQuant (Hadamard + Lloyd-Max) with AWQ (Activation-Aware Weight Quantization) to achieve 93% reduction in quantization error vs standard absmax. This is practically indistinguishable from the full-precision model.

---

🎯 Key Results

---

📊 Quality Evolution

From v1 to v3: 93% reduction in quality loss (0.94 -> 0.06 PPL delta). PolarQuant + AWQ is the key combination.

Cross-Model Results

---

🔬 How It Works

EOQ v3 combines two complementary techniques:

1. AWQ (Activation-Aware Scaling)

Protects important weight channels by pre-scaling them before quantization. Channels that carry more activation energy get higher precision.

2. PolarQuant (Hadamard + Lloyd-Max)

Transforms weight blocks to Gaussian via Hadamard rotation, then applies MSE-optimal Lloyd-Max quantization.

                              AWQ Pre-Scaling
                                    |
                                    v
Original Weights --> Scale Important Channels --> Normalize --> Hadamard Rotate
                                                                      |
                                                                      v
                                                            Lloyd-Max Quantize
                                                                      |
                                                                      v
                                                               Store Codes +
                                                              AWQ Scales +
                                                             Block Norms +
                                                            Centroid Table

Why They Combine Well

• AWQ operates on channels (column-level scaling)
• PolarQuant operates on blocks (128-element sub-vectors)
• They address orthogonal sources of error: AWQ handles channel sensitivity, PolarQuant handles within-block distribution

---

🚀 Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "caiovicentino1/Qwen3.5-9B-EOQ-v3",
    dtype="bfloat16", device_map="auto", trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("caiovicentino1/Qwen3.5-9B-EOQ-v3")

output = model.generate(
    **tokenizer("Write a detailed explanation of neural network quantization:", return_tensors="pt").to("cuda"),
    max_new_tokens=300
)
print(tokenizer.decode(output[0], skip_special_tokens=True))

With torchao INT4 (for maximum speed)

from torchao.quantization import quantize_, Int4WeightOnlyConfig

# After loading EOQ v3 model (already dequanted to BF16):
quantize_(model, Int4WeightOnlyConfig(group_size=128))
# Now runs at 43+ tok/s with 6.5 GB VRAM

---

🔧 Technical Details

Storage Format

{layer_name}.packed     -- bit-packed uint8 quantization codes
{layer_name}.norms      -- fp16 per-block normalization factors
{layer_name}.awq_scales -- fp16 per-channel AWQ importance scales
metadata:
  centroids             -- fp32 Lloyd-Max optimal centroid table (shared)
  bits_per_tensor       -- quantization bits (5 for Q5)

---

📊 Ablation

: Why Both Components Matter

AWQ alone reduces error by 28%. PolarQuant alone reduces error by 80%. Together they reduce error by 94% -- the effects are complementary, not redundant.

---

🔗 Links

• \U0001f4c4 Paper (arXiv) -- PolarQuant: Optimal Gaussian Weight Quantization
• 💻 Code (GitHub) -- Full research codebase
• \U0001f50c vLLM Plugin -- Production inference integration
• 🧊 PolarQuant Q5 -- Without AWQ (simpler, slightly less quality)
• 📊 35B MoE Version -- PPL 5.36, 4.44x compression

---

📖 Citation

@article{vicentino2026polarquant,
  title={PolarQuant: Optimal Gaussian Weight Quantization via Hadamard Rotation for LLM Compression},
  author={Vicentino, Caio},
  journal={arXiv preprint arXiv:2603.7424577},
  year={2026}
}

---

🙏 Acknowledgements

Built with PyTorch, torchao, AWQ methodology from MIT HAN Lab, and the Qwen team's open-weight models.