Repository navigation
Home
Version 2.9.1
Initialize experimental language models from scratch, train them on local data and export supported checkpoints.
Try in your browser · Download the desktop components · README
Open the trial to inspect the interface and choose Check WebGPU adapter. No Python installation is needed. The aim is a quick first look; loading a model or completing a workload depends on the inputs and device.
Native package 2.9.1; browser edition 2.9.
- Model initialization and local corpus-based training controls.
- Live training statistics in a green translucent dashboard.
- A unified progress log that can be saved before closing the page.
- Separate custom WebGPU and installed PyTorch training paths.
A local, inspectable workflow that connects model initialization, training progress and checkpoint export. It supports practical experiments with small language models without making a cloud training service a requirement.
The browser edition packages its interface, JavaScript, CSS and WGSL kernels in one HTML file. This makes online trials easy to open, link to and inspect without a build step or runtime CDN. The file contains the browser engine, not model weights or the native PyTorch backend.
Use a compatible WebGPU browser over HTTPS. Runtime network requests, native hardware monitoring and local-server bridges are disabled. Actual training or evaluation requires supported inputs. The native PyTorch backend requires the full download.
English · Français · 日本語 · 한국어
- Recommended launcher: standard 64-bit CPython 3.12. The installer advertises Python 3.11–3.14, but the bootstrap itself does not enforce that range. Opening it on another version does not establish native compatibility.
-
This Foundry installer reuses an existing compatible PyTorch runtime only. Select it with
--runtime. It does not automatically install a new compute runtime. The selected Python must match its PyTorch wheels and pass the bundled native probe; the reuse path has no independent numeric Python-version gate. A passing older interpreter cannot be inferred from source alone. - The native probe expects PyTorch 2.14.0+rocm10.1.0, HIP and the configured RX 7900 XT/XTX, and checks finite FP32 forward/backward results. The packaged optional dependency helper, if separately invoked, has a narrower installation gate: standard non-free-threaded x64 CPython 3.11–3.14, Windows 11 25H2. Python, GPU drivers and compilers are not installed by the normal Foundry installer. See the official AMD matrix.
Native Windows, WebView2, Radeon and full-model qualification remains pending. The helper’s gate and a successful runtime probe are not proof that every workload is supported.
Initialize experimental models, train with local data and inspect live progress. The green translucent statistics and unified saveable log are retained. Save the log before closing the page.
Use a compatible WebGPU browser on HTTPS or a supported secure localhost origin. The single HTML file contains the custom JavaScript/WGSL engine. Import supported local files. No Python, PyTorch, native service, online dataset download or native GPU telemetry is available; runtime network requests are blocked. GitHub’s source viewer is not an executable trial.
Download every file in the 2.9.1 package together, retaining the supplied layout. Keep all numbered .nxd data parts beside Start_Nexus.py and run python Start_Nexus.py with the supported interpreter. Read the matching guide and installation summary before confirming. Each .nxd part is at most 10,000,000 bytes; each individual distributed file, before installation, is smaller than 20,000,000 bytes. The complete archive, reconstructed application, installed files and downloaded models can be larger. Do not mix releases. Native PyTorch settings and tests belong to the installed edition, not the HTML-only trial.
CPU regression and static/DOM checks are not physical Windows/Radeon/WebGPU validation. Compare only equivalent model, corpus, context, batch, precision and backend conditions. Training/recovery, prompt processing and token generation rates are different measurements. No universal compatibility or maximum speed is guaranteed.
Developer-reported peak training throughput: 32,000 tokens/second on a Radeon RX 7900 XT. Not independently reproduced here, not a sustained-performance guarantee and not chat-generation speed.
Rich Dunbar / Nexus Emerging Technology. Personal use and internal commercial testing and evaluation are permitted for Nexus-owned material. Distribution, modification, production deployment and other commercial use require express written consent. Third-party licenses and previously granted rights remain unchanged. license · evaluation permission (English).