OPENFLUKE · INDEPENDENT AI & ARTIFICIAL-LIFE LAB

Artificial life.
Real possibilities.

Living game worlds, an open-source Go & Python AI engine, and inspectable on-device intelligence built from first principles by Samuel Watson.

AVAILABLE NOW · STEAM · APP STORE · GOOGLE PLAY

Tiny worlds and Sparks you teach by walking the route.

Explore round planets, build bridges between worlds, and train local helpers by demonstration — powered by Welvet and TwinTrace always-on memory across Windows, Linux, macOS, iOS and Android.

NVIDIA CUDAApple MetalApple MLXQualcomm Hexagon HTP NPUVulkan GPUAMD ROCmWebGPUPlan 9 SIMDGo Native EnginePython C-ABI (PyPI)NVIDIA CUDAApple MetalApple MLXQualcomm Hexagon HTP NPUVulkan GPUAMD ROCmWebGPUPlan 9 SIMDGo Native EnginePython C-ABI (PyPI)

THE OPENFLUKE ARCHITECTURE

One lab. Four connected surfaces.

From storage-truth numerical kernels in Go to living planets you can play on desktop and mobile — every layer stays open to inspection.

Characters exploring a tiny green planet in ALife: Becoming
FLAGSHIP GAME · 5 PLATFORMS01 / WORLDS

ALife: Becoming

Explore tiny planets, build hubs and bridges between neighbouring worlds, and teach helpers called Sparks by walking routes yourself. Available on Windows, Linux, macOS, iOS and Android.

GO & PYTHON · v1.7.302 / ENGINE

Welvet Go AI Engine

Open-source Welvet Go AI engine (Welvet Golang AI engine) for local neural network training and inference, now also on PyPI as welvet==1.1.2.

34Storage dtypes
20Quant formats
29Train modes
8+Backends
pip install welvet==1.1.2
IN DEVELOPMENT · 0.1.0+103 / LOCAL APP

Welvet Cross-Platform App

Native application for running and serving 15 approved AI models on hardware you already own — Linux, Windows, macOS, iOS and Android. Round 1 baseline demos confirm working paths.

Fedora-powered OpenFluke Welvet cross-platform AI engine

IN-ENGINE CAPTURES · ALIFE: BECOMING

See the worlds in motion.

Open full 8-scene gallery

HOW IT ALL CONNECTS

Two independent threads, converging into one body of work.

Every prototype, engine rewrite, and benchmark shaped where OpenFluke stands today — built openly in Australia by Samuel Watson.

ENGINE LINEAGE

Paragon → Loom → Welvet Engine (v1.7.3) & App (0.1.0)

Portable Go CPU/SIMD/WebGPU core with opt-in CUDA, Metal, MLX, Vulkan, ROCm, and Qualcomm Hexagon NPU paths.

WORLD LINEAGE

Primecraft · Biocraft · Birdkit · SoulGlitch → ALife OS → ALife: Becoming

Chaos simulation experiments distilled into a physical, playful world where Sparks learn live from your routes.

OPENFLUKE // SYSTEM TOPOLOGYON-DEVICE · EMERGENT
An illustrated artificial world, with connected orbits representing biology, computation, and learning.COMPUTATIONBIOLOGYLEARNING
+THE SPACE BETWEEN LIFE & INTELLIGENCEFIG. 01

OpenFluke & Overview

What is OpenFluke?

OpenFluke is the independent AI and artificial-life research practice of Samuel Watson in Australia. It brings together the Welvet Go AI engine (also called the Welvet Golang AI engine), the Welvet cross-platform app now in development, ALife OS (the earlier prototype), and ALife: Becoming, the game available on Windows, Linux, macOS, iOS and Android.

What is Welvet?

Welvet is an open-source Welvet Go AI engine — a Golang AI engine and machine learning library for neural network training and local inference. Version 1.7.3 runs on CPU tiled, Plan 9 SIMD, and WebGPU by default, with optional NativeCPU, Qualcomm Hexagon HTP (Windows ARM64 and Android), Apple Metal, Apple MLX (forward-only), NVIDIA CUDA, AMD ROCm, and Vulkan backends. The same engine is also on PyPI as welvet==1.1.2 (Python C-ABI). Separately, the Welvet app (0.1.0+1, v0.1 Foundation) is a cross-platform program in development for running and serving models on Linux, Windows, macOS, iOS and Android. Round 1 demos show working paths on real hardware; working means it runs, not that it is fastest or best. Paragon and Loom are earlier engine designs whose lessons continue in Welvet.

What is the Welvet Go AI engine?

The Welvet Go AI engine (Welvet Golang AI engine) is OpenFluke’s open-source AI engine written in Go for training and running neural networks locally. Source: github.com/openfluke/welvet. Product page: openfluke.com/welvet.

Is Welvet a Golang AI engine?

Yes. Welvet is a Welvet Golang AI engine / Welvet Go AI engine: written in Go (Golang), open source, for training and running neural networks locally with optional GPU and NPU acceleration. Source: github.com/openfluke/welvet.

What is the Welvet app?

The Welvet app (version 0.1.0+1, milestone v0.1 Foundation) is OpenFluke’s cross-platform program for running and serving AI models on devices you already own. It is in active development. The On-Device AI Demos series shows shared models across Fedora Linux (CUDA/Vulkan), Windows x86-64 and ARM64, macOS, iOS and Android. Fifteen models are currently approved in the app matrix. Demo results are engineering baselines that improve each round, not claims of maximum performance. See /welvet/app and /welvet/models.

Do I need to understand AI to explore this work?

No. Start with the game and engine overviews. The book and examples are there when you want the technical detail, including source code, recorded outputs, and the limitations documented in the original work.

Welvet Engine & Hardware

Does Welvet support Qualcomm Windows NPU acceleration?

Yes, as of Welvet v1.2 / welvet.qnn 1.0.0. On Windows ARM64 with Snapdragon X (or similar), opt-in BackendQNN plus SyncToQNN / PrepareQNN can run supported layers on Qualcomm Hexagon HTP. It is never the Prefer default. Published large-shape timings are in the Welvet book §81 and w2a go_v_qnn RESULTS.

Does Welvet support Qualcomm Android NPU acceleration?

Yes, as of Welvet v1.7.2 (current engine is v1.7.3). The same opt-in welvet.qnn / Hexagon HTP path runs on Android Qualcomm NPUs, not only Windows Snapdragon X. The measured Android lab matrix (182 rows · 4 fail · 4 timeout) is at files.openfluke.com/1_7/welvet_qnn_all_1794651218607.pdf and on /welvet.

Does Welvet support Python?

Yes, as of Welvet v1.7.3 / PyPI welvet==1.1.2. ctypes bindings over a prebuilt C ABI (create_grid, place_*, train, .entity, list_backends). Linux and macOS arm64 wheels ship the matching native library and optional accelerator siblings. Book chapters §114–117.

Where can I find the Welvet 1.7.2 Android QNN lab results?

On the Welvet page and at files.openfluke.com/1_7/welvet_qnn_all_1794651218607.pdf. It is the public receipt for Android Qualcomm Hexagon HTP under Welvet 1.7.2 (book chapters §79–81 remain the Windows-era writeup).

Does Welvet support Apple Metal acceleration?

Yes, as of Welvet v1.3 / welvet.metal. On Apple silicon macOS, opt-in nativemetal.Mount runs supported layers on the independent Metal GPU library with explicit train, SyncToGo, and JSON Go/C++/Metal routing. It is never the Prefer default and is not Apple Neural Engine. Book chapters §82–88 cover the API and captured Mamba results.

Does Welvet support Apple MLX acceleration?

Yes, as of Welvet v1.7.0 / welvet.mlx 1.0.0. The independent forward-only MLX engine lives at github.com/openfluke/welvet.mlx. On Apple silicon macOS, opt-in BackendMLX / PreferMLX plus SyncToMLX / SyncFromMLX runs supported layers on Metal via libmlx through nativemlx. It is never the Prefer default. No TrainMSE on MLX. Caps reject unsupported layer×dtype. Book chapters §110–113 cover SyncToMLX, caps, Dense/Sequential forward, and go_v_mlx suite pointers.

Does Welvet support NVIDIA CUDA acceleration?

Yes, as of Welvet v1.4 / welvet.cuda 0.79. On Linux x86_64 and Windows amd64 with an NVIDIA GPU, opt-in nativecuda.Mount runs supported layers on the independent CUDA library with explicit train and SyncToGo (two weight copies until you sync). JSON Go/C++/CUDA routing via EnableRouting is Linux amd64. It is never the Prefer default. Book chapters §89–95 cover the API and suite pointers.

Does Welvet support AMD ROCm acceleration?

Yes, as of Welvet v1.5 / welvet.rocm 0.79. The independent AMD engine lives at github.com/openfluke/welvet.rocm. On Linux x86_64 and Windows amd64, opt-in nativerocm.Mount runs supported layers with explicit train and SyncToGo. JSON Go/C++/ROCm routing via EnableRouting is Linux amd64. It is never the Prefer default. Library build/load is validated; GPU compute remains untested until a ROCm-capable HSA device is available. Book chapters §96–102 cover the API and suite pointers.

Does Welvet support Vulkan acceleration?

Yes, as of Welvet v1.6.1 / welvet.vulkan 0.79.1. The independent Vulkan engine lives at github.com/openfluke/welvet.vulkan. On Linux x86_64, opt-in nativevulkan.Mount runs supported layers with explicit train and SyncToGo. JSON Go/C++/Vulkan routing via EnableRouting is Linux amd64. It is never the Prefer default. There is no silent host fallback labeled as Vulkan. Book chapters §103–109 cover the API and suite pointers.

ALife: Becoming & Research

Where can I find the Welvet 1.7 technical research?

On the Welvet page and at the top of the research archive: YouTube youtu.be/f9UBcRNHj4Y, technical research PDF files.openfluke.com/1_7/Welvet 1.7_Technical_Research.pdf, interactive architecture dashboard, feature book v1.7.3 PDF (GitHub release openfluke.github.io v1.7.3), architecture slides (PDF/PPTX), Standard QAT vs Welvet paradigm infographic, continuous on-device training MP3, and the Welvet 1.7.2 Android QNN lab PDF files.openfluke.com/1_7/welvet_qnn_all_1794651218607.pdf. Earlier Welvet 1.3 deep dive remains in the archive.

Where can I find the Welvet 1.3 deep research?

In the research archive (historical): deep-dive PDF at files.openfluke.com/archive/welvet_1_3/Welvet_1.3_Update_Deep_Dive.pdf, audio How_Welvet_Trains_AI_On_Edge_Devices.mp3, and YouTube youtube.com/watch?v=_JHZk1A-J6Q. For the current release, use the Welvet 1.7 technical research pack on /welvet and /research.

Where is the medical AI continual-learning research?

In the research archive: Why medical AI cannot stay frozen. The PDF is files.openfluke.com/archive/polymorphic_play/Medical_AI_Techniques_Research.pdf, the audio briefing is Why_Medical_AI_Cannot_Stay_Frozen.mp3, and the YouTube explainer is youtu.be/dqNkzD_dzPA. It covers locked diagnostic drift, continual learning, federated hospital setups, FDA/TGA adaptive-AI regulation, and OpenFluke’s experimental MORPH controller as a systems testbed.

What is ALife: Becoming?

An exploration and building game where you visit tiny planets, gather resources, connect worlds, and teach Sparks by showing them what to do. It is available on all major platforms: Windows, Linux, macOS, iOS and Android.

What happened to ALife OS?

ALife OS was the prototype. ALife: Becoming is the actual game that grew from it. The older design notebook is preserved as history, rather than a list of completed features.

How do you teach a Spark?

Record a trail by doing the job yourself, then train a Spark to follow the route you demonstrated. The goal is useful automation that stays physical enough to produce surprising outcomes.

Is Becoming available now?

Yes. ALife: Becoming is available everywhere: Windows, Linux and macOS on Steam, iPhone and iPad on the App Store, and Android on Google Play. See openfluke.com/alife for platform links, or email [email protected] with questions.

What happened to the earlier projects?

Primecraft, Biocraft, Birdkit, and SoulGlitch explored different places and configurations for the chaos simulation engine. Those experiments and pivots led through the ALife OS prototype to ALife: Becoming. Paragon and Loom sit on the engine side of that history, leading into Welvet.