AdvancedAI

Bankai: Kilobyte-Scale Patches for 1-Bit LLMs via XOR Adaptation

Bankai is the first post-training adaptation method for true 1-bit large language models. Released on April 2, 2026, it uses bitwise XOR operations on binary weights to create kilobyte-scale patches that modify model behavior with zero inference overhead. Where LoRA adapters weigh ~100 MB each, a Bankai patch compresses to roughly 1 KB — and can be hot-swapped in microseconds.

Bankai project banner showing XOR-based adaptation for 1-bit LLMs
Image credit: Bankai GitHub

The Problem

True 1-bit LLMs like PrismML’s Bonsai represent weights as single bits (0 or 1), enabling an 8B model to fit in just 1.15 GB. But this extreme compression has a cost: traditional adaptation methods like LoRA, fine-tuning, and quantization-aware training all require continuous weights or gradients that binary models lack. Until Bankai, there was no way to adapt a 1-bit model after training.

How XOR Patches Work

Bankai inverts the mechanism of bit-flip attacks — using constructive bit flips for targeted capability improvement instead of sabotage. Each patch is a sparse bitmask specifying which rows (entire neurons of 4,096 bits each) to flip across specific layers and projections. The operation is:

  • Reversible — applying the same XOR patch twice restores the original weights
  • Compact — a typical patch contains 72-93 flips, stored in 864 bytes to 1.1 KB
  • Zero overhead — XOR is a single CPU instruction; no additional inference cost
  • Hot-swappable — microsecond switching between domain-specialized configurations

Results

Validated on Bonsai 8B (the only production-grade true 1-bit LLM), Bankai demonstrated:

  • Scale-guided targeting achieves 3.88× more behavioral impact than random flips
  • A 72-flip arithmetic patch fixes several arithmetic failures while preserving all control probes
  • A generalized patch trained on 60 diverse probes fixes 4 of 17 held-out problems (23.5%) with zero breakage on the 13 already solved
  • No GSM8K degradation — general math reasoning remains intact with patches applied
  • 500K random bit flips (0.009% of MLP weights) cause less than 0.08 perplexity change, revealing massive weight redundancy in 1-bit models

What This Means

Bankai opens a new paradigm for 1-bit model customization. Instead of retraining massive models, developers can create and distribute kilobyte-scale patches for specific capabilities — arithmetic, domain knowledge, safety alignment — and swap them at runtime. As 1-bit models like Bonsai mature, Bankai provides the missing piece: post-deployment adaptation without touching training infrastructure.

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

Related stories

edge computing

PrismML’s 1-Bit Bonsai LLMs: 8B Model in 1.15 GB

On March 31, 2026, PrismML emerged from stealth to announce 1-bit Bonsai — a family of open-weight language models the company calls the first commercially viable 1-bit LLMs. The 8B flagship fits in just 1.15 GB of memory, runs 8× faster than a standard FP16 8B model, and matches its benchmark performance. Backed by $16.25M […]

AI

Google’s TurboQuant Cuts LLM Memory 6x with Zero Accuracy Loss

On March 25, 2026, Google Research introduced TurboQuant — a new compression algorithm that reduces LLM key-value (KV) cache memory by 6x and delivers up to 8x inference speedup on NVIDIA H100 GPUs, all with zero accuracy loss. The algorithm, set to be presented at ICLR 2026, could reshape how organizations deploy large language models […]

Never miss a RITS update

Subscribe to the RSS feed for AI, emerging technology, and digital scholarship news curated for the NYU Shanghai community, or get in touch to suggest a story.