Open Source · github.com/openfluke/loom

The Universal
AI Engine

M-POLY-VTD — a sovereign Deterministic Neural Virtual Machine in Go: 3D volumetric grids, 21 numeric types, and polyglot bindings (welvet) for Python, TypeScript, Flutter/Dart (pub.dev), and WASM. One graph, multiple paths: Plan 9 SIMD CPU, WebGPU / Apple Metal, and vendor NPUs (Intel · Qualcomm Hexagon) — experimental, with measured drift spectra.

v0.83.0 — Experimental Hardware Acceleration Plan 9 SIMD Apple Metal Hexagon NPU
v0.83.0
Experimental Hardware Acceleration
v0.83 — Plan 9 SIMD · Apple Metal · Qualcomm Hexagon · Drift Spectrum
Plan 9 SIMD — raw AVX2 / NEON across the engine · no CGO · no CUDA / OpenBLAS stack
Apple Metal — zero-copy unified memory on Apple Silicon · Lucy [13]
Hexagon NPU — Qualcomm QNN plugin · Drift Spectrum (EXACT · INDUS · LOWBIT · H-DRIFT)
Lucy seven-layer.entity save/reload · SC/MC train · multi-dtype determinism
v0.83 · Experimental Hardware Acceleration

Watch Loom hit SIMD, Metal & Hexagon

The v0.83 release story: Plan 9 SIMD without CGO, Apple Metal zero-copy on Apple Silicon, and Qualcomm Hexagon NPU offload with the Drift Spectrum — sovereign DNVM on divergent hardware.

AI Deep Research

Independent AI Analysis of Loom

Comparative research on M-POLY-VTD vs PyTorch and JAX — plus the full engine reference on this site, synced from loom/docs.

Architecture & research
3D grids, target propagation, DNVM
Start with the overview — volumetric dispatch, WeightStore morphing, step mesh, transformers, v0.83 Experimental Hardware Acceleration (SIMD · Metal · Hexagon), v0.81 Accelerator Bridge, and v0.80 Native Ship (.entity checkpoints).
🧊
AI that thinks in 3D

Most AI frameworks process data in a straight line, like an assembly line. Loom uses a three-dimensional grid — more like how your brain's neurons actually connect, jumping across regions rather than always going layer by layer.

💾
Fits AI on a USB stick

Loom can compress AI models by up to 98.4%. A model that normally takes gigabytes of storage can shrink to a fraction — small enough to run on a phone or an old laptop with no internet required.

🧬
Learns like biology, not math

Traditional AI learning requires freezing everything to calculate one massive equation. Loom's Target Propagation lets each part of the network learn independently — more like how neurons fire and strengthen in a real brain.

Read the Full Technical Breakdown
For Non-Technical People

What is Loom, exactly?

"Think of Loom like SQLite — but for AI."

SQLite is a tiny database that runs inside your app with no server needed. Loom is the same idea for neural networks: a self-contained engine you can drop into any project, on any device, with no cloud account, no GPU server, no complicated setup.

🧠
Train it like a brain

A neural network learns by seeing examples — like showing a child thousands of pictures of cats until they know what a cat is. Loom provides all the tools to build and teach these networks.

📦
Pack it anywhere

Once trained, your model is a tiny file. Drop it into your Python script, your phone app, your website, or a game engine. Loom runs it everywhere with the exact same output.

🔒
No cloud needed

Unlike ChatGPT or other AI services, Loom runs 100% locally on your device. Your data never leaves your machine. Perfect for privacy-sensitive apps or offline use.

CPU · SIMD · GPU · NPU

Loom CPU runs on every OS and binding — with Plan 9 SIMD (AVX2 / NEON) from v0.82/v0.83. WebGPU and Apple Metal hit the GPU; vendor plugins offload layers to Intel OpenVINO and Qualcomm Hexagon NPUs (experimental), with measured Drift Spectrum parity. Accelerators guide →

🌍
Every language

Python developer? pip install welvet. JavaScript? npm install @openfluke/welvet. Flutter? welvet: ^0.80.4 on pub.dev — SoulGlitch and loom-flutter-quickstart prove it ships. Go, C, C#, Rust? There are bindings for all of them. One model, every language.

🎯
Deterministic on CPU & GPU

Loom's Deterministic Neural Virtual Machine (DNVM) delivers bit-identical behaviour across Apple Silicon, x86, WebGPU, and language bindings. Lucy and SoulGlitch depend on this for reproducible local inference.

🧬
Evolution built in

Loom includes a full NEAT evolution engine — models can mutate and breed like living organisms. This powers SoulGlitch's creature evolution system.

Lucy & SoulGlitch

Supported Hugging Face models

Approved checkpoints share the same list in loom/lucy and SoulGlitch—download once, run offline via welvet.

SmolLM2

135M · 360M · 1.7B Instruct — mobile to server brains

Qwen3

0.6B · 1.7B · 4B — GPU-friendly chat models

BitNet b1.58

microsoft/bitnet-b1.58-2B-4T — packed ternary CPU path (v0.78+ infer · v0.79+ native save/reload · v0.80 .entity ship)

Plus custom Loom/poly networks (training, NEAT, DNA) with no HF download.

Flutter

Run Loom in Flutter apps

welvet on pub.dev is an FFI plugin — same C-ABI as Python and npm. Desktop natives ship in the package; SoulGlitch proves production mobile.

📦
pub.dev install

welvet: ^0.80.4 in pubspec.yaml. No monorepo path. loomAvailable + loomLib on first FFI call. Package page →

🧪
loom-flutter-quickstart

30+ interactive CPU demos: dense/MHA forward, train, mesh step, DNA, checkpoints, seven-layer suites. Clone & run →

📖
Docs & snippets

Forward, morph, LoomTrain, backward, .entity wire, LLM exports. Flutter guide → · Deployment →

Get Started

Install in 30 seconds

Pick your language and paste the command. No account required.

Python
Node.js
Flutter
Go
WebAssembly
$ pip install welvet

Ships with precompiled native libraries for Windows, Linux, macOS, iOS, and Android. Zero Python dependencies. PyPI page →

$ npm install @openfluke/welvet

Works in Node.js and browsers via WebAssembly. npm page →

welvet: ^0.80.4

Add to pubspec.yaml — FFI plugin with desktop natives on pub.dev. pub.dev → · Flutter docs → · Example app →

import 'package:welvet/loom_ffi.dart';

loomLib.configureTrainingMode(handle, 2);
final out = loomParseFloatArray(
  loomLib.forwardPolymorphic(handle, input, [1, 16]),
);
$ go get github.com/openfluke/loom/poly

Pure Go module. No CGO. Works with standard go build. Quick reference → · Source →

Download main.wasm from the releases page Download

6.9 MB WASM bundle. Drop into any web page and run Loom in the browser. All releases →

Platform Support

Runs everywhere

Prebuilt native libraries for every major platform — just download and go.

Windows
x86-64, ARM64
Linux
x86-64, ARM64, ARM v7, x86
macOS
x86-64, ARM64 (M-series), Universal
Android
ARM64, x86-64
iOS
ARM64, Simulator, XCFramework
WebAssembly
Browser + Node.js
Vendor NPU
Intel · Qualcomm Hexagon · experimental · Drift Spectrum
Apple Metal
MPSGraph · macOS Apple Silicon · Lucy [13]
Plan 9 SIMD
AVX2 · NEON · no CGO
WebGPU
Forward + Backward pass, 17×–65× speedup
Flutter / Dart
welvet on pub.dev · desktop natives · SoulGlitch + quickstart demos
PyPI
welvet — zero dependencies
For Developers

What's under the hood

Loom isn't just a wrapper around PyTorch. It's a ground-up engine built for portability and precision.

All major layer types

Dense, MHA, SwiGLU, RMSNorm, LayerNorm, CNN 1D/2D/3D, Transposed Conv, RNN, LSTM, Embedding, KMeans, Softmax, Parallel, Sequential, Residual.

21 numeric types

float64 all the way down to binary (1-bit), including fp8, fp4, int4, and ternary. Choose precision vs. model size at runtime.

NEAT evolution + DNA

A full neuroevolution engine with mutation, crossover, and fitness selection. Models have a "DNA" signature for reproducible evolution.

98.4% compression

Native bit-packed serialization shrinks model files by 98.4% compared to raw float storage. Plus SafeTensors support for HuggingFace compatibility.

Target propagation

An alternative to backpropagation where each layer is given a direct target. More biologically plausible and works for non-differentiable layers.

Step mesh engine

Clock-cycle 3D grid with double-buffered layers, spatial remote links, BPTT, and neural target propagation — online learning without a rigid layer stack.

BitNet & low-bit CPU

BitNet b1.58–style checkpoints with packed ternary linear layers. Lucy pulls from Hugging Face; welvet C-ABI exposes CPU inference paths.

Operation mesh

Donate Compute (LAN TCP model sharing), TANHI UDP layer telemetry for SoulGlitch HUD, tiled forward/backward, and Qwen3-family HF ingest.

Full documentation Deployment guide BitNet CPU
Watch It Work

See Loom In Action

Live demos — TANHI telemetry to SoulGlitch, WebGPU benchmarks, and 3D visualization.

Loom × SoulGlitch · live
TANHI × Regional Mix — models on your PC, view on your phone

Watch Loom AI models run in real time on a regional_mix harness (Dense, MHA, SwiGLU, RNN, LSTM with remote links across 3D topologies). Execution streams over UDP as TANHI telemetry into SoulGlitch on your local phone — a spatial, time-scrubbable trace instead of numbers in a terminal.

TANHI docs → · YouTube →
Performance Benchmark
Forget Llama.cpp: WebGPU Inference in Pure Go

SmolLM2-135M benchmarks: 68 tok/s on RTX 1650 Super, 143 tok/s on Linux i5, 229 tok/s on Mac M4. Zero CGO. FlashPoly Tiling. Bit-level deterministic across OS boundaries.

Visualization
Loom: Visualizing 3D Neural Networks in Real-Time

Watch the AI "think" in real-time. Stepping mode, 3D grid topology, Zig-Zag and Starburst routing patterns — the black box, opened.

Android · Airplane Mode
Offline LLM Inference on Android via Loom AI

Loom v0.0.8 running 100% locally on Android — device locked in Airplane Mode throughout. Zero cloud dependency. Pure on-device compute from first principles in Go.

Open Source Tool
NeuralWave: 3D Neural Network Visualization & Weight Analysis

Real-time model discovery from HuggingFace, interactive 3D layer inspection, attention head visualization. Built on Loom + Go backend + Three.js.

Star Loom on GitHub

Loom is free, open-source, and built in the open. Stars help others find it and fuel continued development.

Star   Fork