NEURAX

An environment to design, simulate, and optimize neural architectures — before you spend a single GPU-hour finding out what they cost.

The usual way to find out an architecture doesn't fit is to launch the training run and watch it OOM six hours in or finish, and cost three times what you budgeted, because nothing checked before you committed the GPU-hours. NEURAX checks first. Point its analytical engine at an architecture a full LLM or a small model built for your own dataset and it hands back memory, speed, and training cost in under 50 milliseconds. No GPU, no training run, no waiting to find out.

_{A 28B GQA/SwiGLU transformer, loaded and analyzed — parameters, FLOPs, VRAM, and roofline position, all computed the moment it was built.}

Linux or macOS — one command:

curl -fsSL https://raw.githubusercontent.com/rustnew/NEURAX/main/install.sh | sh

Then type neurax, or open NEURAX from your applications menu. That's the entire install — no sudo, nothing written outside your home directory, nothing to sign up for.

_{The command above, actually run — download, install, and NEURAX ready to open, in one pass.}

Runs on Debian, Ubuntu, Kali, Arch, Fedora, and effectively any Linux, plus macOS (Intel and Apple silicon).

Windows — no one-line installer (install.sh is a POSIX shell script; Windows' own shells don't run it):

1. Go to Releases.
2. Under Assets, download NEURAX_<version>_x64-setup.exe (or NEURAX_<version>_x64_en-US.msi, if your organization prefers MSI installers).
3. Run the downloaded file. Windows 11 already has the WebView2 runtime NEURAX renders through; the installer adds it on Windows 10 if it's missing.
4. Launch NEURAX from the Start menu.

Same install as Linux and macOS otherwise: no account to create, no data leaves the machine, and the whole environment runs locally, in the app itself.

Not sure yet? Read the install script before running it — one file, plain shell: install.sh.

Drag blocks onto a canvas, or load one of 64 real reference architectures across the 8 architecture families the compiler fully supports — LLM, diffusion, MoE, and more — and get:

• Will it fit? Peak VRAM, activations, optimizer state — to the byte.
• What will it cost? Training time, dollars, energy, CO₂ — on the hardware you actually have.
• Where's the bottleneck? FLOPs by operation, per layer, memory- or compute-bound.
• Will it behave? Inference stability and hallucination risk, before you serve a token.

Every number is computed from the architecture you built, live — not a lookup table.

_{Left: import any public model straight from HuggingFace — paste an ID or a config.json, nothing uploaded. Right: the same environment, on a diffusion architecture — not just LLMs.}

This is what the opening line means, in practice — a 512-expert, 92-layer design pushed deliberately past a single GPU, caught before any training run:

_{21.5 TB of peak VRAM against an 80 GB card — NEURAX says so instantly, not partway through a training run that was never going to finish.}

_{Time Machine also tracks what training that model would actually be obligated to — EU AI Act thresholds, carbon reporting, verified against real, dated regulation text, not a generic checklist.}

_{Change an architecture decision and see exactly what moved — every metric, as a percentage, not just a new number to compare by eye.}

This isn't only for people who already have a research cluster. A small, specialized model tuned to your own dataset gets the same precise answer as a 70B one. That's the common case NEURAX is built for.

• No account to create — a local profile is generated automatically, kept on your machine.
• No API key leaves your browser — the AI copilot (bring your own key: OpenAI, Anthropic, Gemini, Mistral, Fireworks, DeepSeek, GLM, or any custom endpoint) talks straight to your chosen provider.
• No project ever uploaded anywhere. The whole environment runs on your machine; that's the whole design.

Every reference model is checked against its real published size by an automated test — not eyeballed once:

Four of these were wrong before that test existed — one by +122%. That's why the test exists, and why it still runs on every change.

The one-liner above is the reference way to run NEURAX — the desktop app, with the compiler embedded. For everything else — Windows and web builds, Docker, running it as a Rust library, the full architecture, the API reference — see the documentation.

Contributing · Changelog · Security · License: MIT

Built by Fossouo.