nikshepsvn

nikshepsvn / bankai

Public

Ultra-Sparse Adaptation of 1-Bit LLMs via XOR Patches

github.comnikshepsvnbankai 1-bit apple-silicon binary-neural-networks bonsai llm
10
2
89% credibility
Found Apr 02, 2026 at 10 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Bankai creates tiny, reversible tweaks for ultra-compressed AI language models to improve targeted skills like math or coding while keeping overall performance intact.

How It Works

1
🔍 Discover Bankai

You hear about a clever tool that helps make super-small AI helpers better at specific things like solving math puzzles.

2
đź“Ą Get the tool ready

You grab the easy-to-use kit and set up your tiny AI brain on your computer.

3
Pick a skill to boost
âž•
Boost math

Make it great at addition, multiplication, and more.

đź’»
Boost code

Help it understand simple programming better.

🧠
Boost knowledge

Sharpen everyday facts without losing other smarts.

4
⚡ Find the perfect tweak

The tool smartly searches and creates a tiny, safe change that boosts your chosen skill.

5
🧩 Apply the tweak

You add this lightweight magic to your AI, and it feels instantly sharper.

6
🧪 Test the results

Ask your AI tough questions and watch it nail them right, while still good at everything else.

🎉 Smarter AI unlocked

Your compact AI now shines at your favorite tasks, all with a simple, reversible upgrade.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 10 to 10 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is bankai?

Bankai lets you create ultra-sparse XOR patches for 1-bit LLMs like Bonsai-8B, adapting their behavior on specific tasks like math or code without full retraining. Using Python and MLX-LM, you run a CLI search to optimize patches against probe prompts—flipping entire rows in MLP weights—then apply them reversibly in one bitwise operation. Inspired by bankai bleach and bankai ichigo's ultimate power (bankai japanisch for "final release"), it delivers KB-sized files that boost logit gaps on targets while penalizing control degradation.

Why is it gaining traction?

Unlike LoRA or full fine-tuning, bankai patches are ultra-sparse via XOR—reversible, zero-overhead to apply, and tiny enough for on-device tweaks to 1-bit LLMs. Developers dig the greedy search CLI that screens candidates fast on Apple Silicon, yielding patches with proven generalization on unseen calculus or prime probes. It's a fresh take on 1-bit adaptation, blending bankai senbonzakura kageyoshi precision with practical eval tools.

Who should use this?

ML engineers deploying 1-bit LLMs on MacBooks for edge inference, needing quick math or knowledge fixes without cloud training. Researchers probing sparse LLM edits, or devs customizing bankai architekten-style models for domain-specific prompts like GSM8K safety checks.

Verdict

Grab it if you're experimenting with ultra-sparse 1-bit LLM patches—experiments show real generalization wins—but with 10 stars and 0.9% credibility, it's raw; expect to tweak probes yourself as docs are README-only.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.