BWLA: Breaking the Barrier of W1AX Post-Training Quantization for LLMs

> 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 Big to Run" Problem

Imagine running an LLM locally on a phone:

• LLaMA-3 8B model: 16GB (FP16)
• Phone memory: 8GB
• It simply won't fit.

Existing quantization approaches:

• Weight-only INT4 quantization: 4GB
• But activations remain 16-bit
• Activation caching during inference is still large
• No true end-to-end speedup

The real bottleneck: activations.

The Challenge of W1A? Quantization

W1: binarize weights to 1 bit

• Each weight is only +1 or -1
• 32x weight storage compression
• Simplified compute (additions only, no multiplications)

AX: quantize activations to X bits

• Activations have heavy-tailed distributions
• Extreme outliers destroy quantization accuracy
• Existing methods cannot handle this

Why are activations harder to quantize than weights?

Weights are fixed after training with relatively uniform distributions, so it's easy to find good quantization ranges. Activations change with every input, their distributions are dynamic, and they contain extreme outliers — leading to large accuracy loss after quantization.

BWLA: Breaking the W1AX Barrier

The paper proposes BWLA (Binarized Weights and Low-bit Activations):

> The first method to achieve W1A8 (1-bit weights + 8-bit activations) post-training quantization, enabling true end-to-end low-bit inference.

Technical approach:

1. Weight binarization (W1): Map weights to +1/-1 using post-training quantization (PTQ) — no retraining required. 2. Low-bit activations (A8): The key breakthrough is handling activation outliers — identify and isolate extreme values, quantize the normal range to 8 bits, and handle outliers with higher precision or special treatment. 3. End-to-end optimization: Quantize not only weights but also intermediate activations, reducing memory footprint and bandwidth for real hardware acceleration. 4. Accuracy preservation: A carefully designed quantization strategy keeps accuracy loss minimal, validated across multiple tasks.

Results:

• Weight storage: 32x compression
• Activation storage: 2x compression
• Compute: substantially reduced
• Accuracy: close to the original model

It's like turning a hardcover book into a paperback: same content, much smaller size, readable anywhere.

Why End-to-End Quantization Matters

Problems with weight-only quantization:

• Inference bottleneck: weights shrink but activations stay large; KV cache consumes significant memory, limiting long-context handling
• Compute bottleneck: weight reads decrease, but activations are still computed at full precision — total compute doesn't drop

BWLA's advantages:

• Full compression: weights + activations + KV cache, all low-bit — genuine end-to-end speedup
• Hardware-friendly: 1-bit weights need only adders; 8-bit activations use standard INT8 ops supported by existing AI chips
• Flexible deployment: edge devices, mobile, low-power scenarios

Takeaways for Practitioners

If you're optimizing model deployment or quantization, ask:

1. Does my quantization focus only on weights while ignoring activations? 2. Are activation outliers the bottleneck? 3. Is end-to-end low-bit inference necessary for my use case? 4. How do I preserve accuracy under extreme compression?

BWLA reminds us that real model compression isn't "good enough" — it's "doing the best under extreme constraints." With 1-bit weights and 8-bit activations, an LLM transforms from a data-center giant into something that fits in your pocket. On the road to democratized AI, BWLA is a meaningful milestone.

In the world of bits, sometimes 1 bit is enough.