Paper: BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs Authors: Zhixiong Zhao, Zukang Xu, Dawei Yang arXiv: 2605.00422 | 2026-04-29
The "Model Too Large to Run" Dilemma
Imagine running an LLM locally on your phone:
- LLaMA-3 8B model: 16GB (FP16)
- Phone memory: 8GB
- It simply doesn't fit
- Weight-only INT4: 4GB
- But activations remain 16-bit
- Activation caching during inference still costs a lot
- No true end-to-end speedup
- 32x storage compression per weight
- Multiplications become additions
- Activations have heavy-tailed distributions
- Extreme outliers destroy quantization accuracy
- Existing methods fail to handle them
- Weight storage: 32x compression
- Activation storage: 2x compression
- Compute: substantially reduced
- Accuracy: close to the original model
- Weights shrink, but activations stay large
- KV cache consumes huge memory; long contexts become impractical
- Activation compute remains full-precision, so compute doesn't drop
- Compresses weights, activations, and KV cache together for real end-to-end speedup
- Hardware-friendly: 1-bit weights need only adders; 8-bit activations map to standard INT8 units supported by existing AI chips
- Flexible deployment: edge devices, mobile, low-power scenarios
Existing quantization approaches:
The real bottleneck: activations.
Why W1AX Quantization Is Hard
W1 — binarizing weights to +1/-1:
AX — quantizing activations to X bits:
Why are activations harder than weights?
Weights are fixed after training with relatively uniform distributions, making it easy to find good quantization ranges. Activations vary with every input, their distributions shift dynamically, and they contain extreme outliers—so quantization error is much larger.
BWLA: Breaking the W1AX Barrier
BWLA (Binarized Weights and Low-bit Activations) claims to be:
> The first post-training quantization method achieving W1A8 (1-bit weights + 8-bit activations), enabling true end-to-end low-bit inference.
Key techniques:
1. Weight binarization (W1): Map weights to +1/-1 via post-training quantization—no retraining required. 2. Low-bit activations (A8): The core breakthrough—identify and isolate activation outliers, quantize the normal range to 8 bits, and handle outliers with higher precision or special treatment. 3. End-to-end optimization: Quantize not just weights but intermediate activations, cutting memory footprint and bandwidth for genuine hardware acceleration. 4. Accuracy preservation: Careful quantization strategies keep accuracy loss minimal, validated across multiple tasks.
Results:
It's like turning a hardcover book into a pocket edition—same content, far smaller size, readable anywhere.
Why End-to-End Quantization Matters
Weight-only quantization limitations:
BWLA advantages:
Takeaways
If you're optimizing model deployment or quantization, ask yourself:
1. Does my quantization focus only on weights and ignore activations? 2. Are activation outliers the real quantization bottleneck? 3. Is end-to-end low-bit inference necessary for my use case? 4. How do I preserve accuracy under extreme compression?
BWLA is a reminder that true model compression isn't "good enough"—it's doing the best under extreme constraints. When weights are 1 bit and activations are 8 bits, an LLM transforms from a data-center giant into a pocket-sized companion. On the path to democratizing AI, BWLA marks an important milestone.
Sometimes in the world of bits, 1 bit is enough.