0xbitnet

0xBitNet

Run Microsoft BitNet b1.58 ternary LLMs with WebGPU — in browsers and native apps.

Live Chat Demo · TL;DR Widget Demo · Getting Started · API Reference

---

0xBitNet runs BitNet b1.58 ternary LLMs on WebGPU. Custom WGSL compute kernels handle the ternary matrix operations, with bindings for TypeScript, Rust, and Python. Works in browsers, Node.js, and native apps.

Highlights

• Pure WebGPU — Custom WGSL kernels for ternary matrix operations (no WASM, no server)
• Multi-language — TypeScript (0xbitnet), Rust (oxbitnet), Python (oxbitnet), Swift (OxBitNet), Java/Android (oxbitnet-java), C (oxbitnet-ffi)
• Cross-platform — Browsers, Node.js, Deno, native apps via wgpu
• Chat templates — Built-in LLaMA 3 chat message formatting
• Automatic caching — IndexedDB (browser) / disk cache (native)
• Streaming — Token-by-token output via async generators / streams / callbacks

Quick Start

TypeScript / JavaScript

npm install 0xbitnet

import { BitNet } from "0xbitnet";

const model = await BitNet.load(
  "https://huggingface.co/microsoft/bitnet-b1.58-2B-4T-gguf/resolve/main/ggml-model-i2_s.gguf",
  { onProgress: (p) => console.log(`${p.phase}: ${(p.fraction * 100).toFixed(1)}%`) }
);

for await (const token of model.generate("The meaning of life is")) {
  process.stdout.write(token);
}

model.dispose();

Rust

cargo add oxbitnet

use oxbitnet::BitNet;
use futures::StreamExt;

let mut model = BitNet::load("model.gguf", Default::default()).await?;

let mut stream = model.generate("Hello!", Default::default());
while let Some(token) = stream.next().await {
    print!("{token}");
}

model.dispose();

Python

pip install oxbitnet

from oxbitnet import BitNet

model = BitNet.load_sync("model.gguf")

model.chat(
    [("system", "You are a helpful assistant."), ("user", "Hello!")],
    on_token=lambda t: print(t, end="", flush=True),
    temperature=0.7,
)

model.dispose()

Swift

import OxBitNet

let model = try await BitNet.load("model.gguf")

for try await token in model.chat([.user("Hello!")], options: .init(temperature: 0.7)) {
    print(token, terminator: "")
}

model.dispose()

Java

import io.github.m96chan.oxbitnet.*;
import java.util.List;

try (BitNet model = BitNet.loadSync("model.gguf")) {
    model.chat(
        List.of(new ChatMessage("user", "Hello!")),
        token -> {
            System.out.print(token);
            return true;
        },
        new GenerateOptions().temperature(0.7f)
    );
}

C / FFI

#include "oxbitnet.h"

static int32_t on_token(const char *token, uintptr_t len, void *userdata) {
    fwrite(token, 1, len, stdout);
    return 0; /* 0 = continue, non-zero = stop */
}

int main(void) {
    OxBitNet *model = oxbitnet_load("model.gguf", NULL);

    OxBitNetChatMessage messages[] = {
        { .role = "user", .content = "Hello!" },
    };
    OxBitNetGenerateOptions opts = oxbitnet_default_generate_options();

    oxbitnet_chat(model, messages, 1, &opts, on_token, NULL);
    oxbitnet_free(model);
}

Chat Messages (TypeScript)

const messages = [
  { role: "system" as const, content: "You are a helpful assistant." },
  { role: "user" as const, content: "Explain quantum computing in one sentence." },
];

for await (const token of model.generate(messages, { maxTokens: 128, temperature: 0.7 })) {
  process.stdout.write(token);
}

Supported Models

Any I2_S GGUF model with a compatible architecture should work — see Model Compatibility for details.

Install

API Overview

TypeScript

Rust

Python

Swift

Java

C / FFI

Platform Support

0xBitNet runs on any platform with a WebGPU implementation:

Browsers:

• Chrome / Edge 113+ (recommended)
• Firefox Nightly (behind flag)
• Safari 18+

Native (Rust / Python):

• Uses wgpu — Vulkan, Metal, DX12 backends automatically
• No browser or WebGPU runtime needed

Native (Node.js / Deno):

• Deno (built-in WebGPU)
• Node.js with webgpu npm package (Dawn bindings) — see Node.js CLI example
• Any runtime exposing the WebGPU API (e.g., wgpu-native, Electron)

A dedicated GPU with sufficient VRAM is required (see Supported Models for estimates).

Examples

Web Chat

A WebGPU-powered chat application. Downloads the model on first visit, then runs LLM chat completely on-device — no backend needed.

TL;DR Widget

An offline-ready summarization widget. Provides LLM-powered TL;DR without any network dependency.

Node.js CLI

Run BitNet from the command line using Node.js and the webgpu npm package (Dawn bindings). Interactive chat with streaming output and tok/s metrics.

cd examples/node-cli
npm install && npm start

Rust CLI

Interactive chat using native wgpu.

cd packages/rust
cargo run --example chat --release

Python CLI

Interactive chat via Python bindings.

pip install oxbitnet
python packages/rust/crates/oxbitnet-python/examples/chat.py

Swift CLI

Minimal Swift chat example wrapping the C FFI layer.

cd packages/rust
cargo build -p oxbitnet-ffi --release
cd crates/oxbitnet-swift
swift run -Xlinker -L../../../../target/release Chat model.gguf "Hello!"

Java CLI

Minimal Java chat example using JNI bindings.

cd packages/rust
cargo build -p oxbitnet-java --release
cd crates/oxbitnet-java/examples
javac -cp ../java/src/main/java:. Chat.java
java -Djava.library.path=../../../../target/release -cp ../java/src/main/java:. Chat model.gguf "Hello!"

C CLI

Minimal C example using the FFI bindings.

cd packages/rust
cargo build -p oxbitnet-ffi --release
gcc crates/oxbitnet-ffi/examples/chat.c -Icrates/oxbitnet-ffi -Ltarget/release -loxbitnet_ffi -o chat
LD_LIBRARY_PATH=target/release ./chat model.gguf "Hello!"

Architecture

0xbitnet/
├── packages/
│   ├── core/               # WGSL kernels + TypeScript API (npm: 0xbitnet)
│   │   └── src/
│   │       ├── gpu/        # WebGPU device init, buffer pool
│   │       ├── model/      # GGUF parser, weight loader, config
│   │       ├── nn/         # Transformer layers, attention, BitLinear
│   │       ├── shaders/    # 12 WGSL compute shaders (shared with Rust)
│   │       └── tokenizer/  # BPE tokenizer, chat templates
│   └── rust/               # Rust + Python bindings
│       └── crates/
│           ├── oxbitnet/           # Rust library (crates.io: oxbitnet)
│           ├── oxbitnet-python/    # Python bindings via PyO3 (PyPI: oxbitnet)
│           ├── oxbitnet-swift/     # Swift bindings via C FFI (SPM package)
│           ├── oxbitnet-java/      # Java/JNI bindings (Android-ready)
│           └── oxbitnet-ffi/       # C FFI bindings (cdylib + staticlib)
├── examples/
│   ├── web-chat/           # Chat app demo (Vite)
│   ├── tl-dr-widget/       # Offline TL;DR widget demo (Vite)
│   └── node-cli/           # Node.js CLI using Dawn WebGPU bindings
└── docs/

See Architecture for data flow and internals.

Prerequisites

• TypeScript/JS: Node.js 18+, a WebGPU-capable environment
• Rust: Rust 1.75+, a Vulkan/Metal/DX12-capable GPU
• Swift: Swift 5.9+, a Vulkan/Metal/DX12-capable GPU
• Java: JDK 17+, a Vulkan/Metal/DX12-capable GPU
• Python: Python 3.9+, pip install oxbitnet

Contributing

Contributions are welcome! Whether it's a bug report, feature request, or pull request — all input is appreciated.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

Please see CONTRIBUTING.md for detailed guidelines.

License

MIT