Termux-BitNet

High-Performance 1.58-bit On-Device LLM Inference SIMD Engine for ARM64 & Android Termux

1-Line Quick Installation

Install the official package directly into your runtime:

pip install termux-bitnet
# or:
npm install termux-bitnet

The Engineering Challenge

Standard LLMs (FP16/INT8) cause severe memory bottleneck, excessive battery drain, and thermal throttling on mobile ARM64 devices.

The Architectural Breakthrough

Features ARM64 NEON DotProd SIMD kernels executing ternary weights {-1, 0, +1} with sub-15ms token latency on smartphone processors.

Key Capabilities & Built-in Hardening

Deterministic 0-Drift Output

Bit-exact floating-point precision verified across heterogeneous ARM64 & WebGPU hardware.

Zero Cloud Egress Architecture

Operates 100% on the local client without external network telemetry leaks.

Memory Leakage Protection

Weakref lifetime management preventing GPU VRAM / system RAM leaks.

Supported Compute Kernels & Operations

Subsystem Category	Operations & Kernels	Status
Compute Kernel	Hardware Native Accelerated Kernels	Production
Memory Subsystem	Zero-Copy Ring Buffers & Weakref GC	Production
Runtime Compatibility	Android Termux Bionic / Browser WebGPU	Production

Canonical Usage Example

import termux_bitnet as tbn

engine = tbn.BitNetEngine(model_path="bitnet_b1_58_3b.tbn", threads=4)
for token in engine.generate_stream("What is sovereign on-device AI?"):
    print(token, end="", flush=True)

import { BitNetEngine } from "termux-bitnet";

const engine = new BitNetEngine({ modelPath: "bitnet_b1_58_3b.tbn", threads: 4 });
const stream = await engine.generateStream("Hello ARM64 BitNet!");
for await (const token of stream) { process.stdout.write(token); }