Version 25
A local AI workspace for importing supported language models and running chat through the custom WebGPU or Nexus.cpp inference engine.
Open the browser trial · Download the full application · Installation guide · Wiki
Open the trial, choose your GPU profile and memory capacity in Models & settings, then select a local GGUF model and load it for chat. It uses the existing custom WebGPU inference engine, not a simulated conversation. A working WebGPU adapter, sufficient memory and a compatible model are required. No model is bundled. The loader checks the model's architecture, tensors, quantization and tokenizer; a filename or hardware profile alone does not establish compatibility.
The single HTML file embeds the interface, inference code and tokenizer/template library so an evaluator can open the application without Python, an installer or a build step. Model selection and loading are separate and can take longer. The trial does not use a local server, upload model data or provide a CPU fallback. Native Nexus.cpp, online services, voice, camera and teacher features require other components and are not included in this browser trial.
The full local application remains a separate download. Its existing Candidate25 directory and package hashes are unchanged. Public version: 25; browser packaging revision: v25-browser-r1.
- Imports local models and checks the selected engine's architecture, tokenizer and quantization support.
- Provides local chat, explicit hardware profiles and capability-aware controls.
- Supports cooperative Stop, New Chat reset, session-scoped prefix caching and min-p/repetition controls through native protocol 3.
- Records bounded diagnostics, rotating logs and registry-recovery information.
- Keeps model admission, loaded weights, server health and chat readiness separate.
Nexus AI brings custom browser and native inference paths into one local workspace. Its contribution is the integration of model checks, controllable conversation state and inspectable diagnostics, rather than a claim that every model or GPU is supported.
The engines have different capabilities. An existing native executable must be rebuilt for protocol 3 features. Native batch remains one token per step; native Flash Attention and reduced-precision KV are not implemented in this release.
- Download every file in releases/Candidate25, preserving the supplied layout.
- Keep both numbered data files with
Nexus_AI_Manifest.json, the installer, GUI, artwork and dependency helper. - Run Install_Nexus_AI.py with CPython 3.10 or later. The recorded Linux checks used Python 3.12.14. Review the installation plan before confirming. The command-line option is
python Install_Nexus_AI.py --cli;--no-launchinstalls without opening the app.
The optional desktop window uses pywebview>=5,<7. Browser GPU work needs a compatible WebGPU browser. Nexus.cpp needs a compatible built executable and the appropriate compiler, HIP and driver setup for Radeon acceleration. Nexus AI uses its own inference engines, not a PyTorch backend.
Each distributed data part is at most 10,000,000 bytes; every individual distribution file is below 20,000,000 bytes. The installed application and downloaded model weights can be larger.
Mistral 7B Instruct v0.3 is a useful first chat model to consider when the chosen engine supports its exact format. The official model card distributes Safetensors; the GGUF loader needs a trustworthy, compatible GGUF conversion instead. Check provenance, license and a successful load before using it.
RX 7900 XT and XTX profiles use their nominal 20 GB and 24 GB VRAM capacities. Available memory is lower once other processes, runtime allocations and context/KV cache are included. These are user-selected nominal memory budgets, not measurements of free VRAM. Selecting a profile is not a benchmark result. Full Mistral 7B operation on Windows/Radeon has not been qualified by the recorded tests.
The standalone packaging has passed scripted no-network boot checks. GPU generation and full-model quality/performance qualification are separate from those checks; no universal model support or tokens-per-second result is claimed here. Any load-time finite/nonconstant-logit health check is a basic readiness check, not an evaluation of answer quality.
The package includes producer regression results, focused integration checks, readiness boundaries and a Word guide. Focused checks covered installer/recovery behavior, native conversation controls, sampler/prefill behavior and all 408 payload hashes.
These are scoped software checks, not enterprise certification, physical Windows/Radeon qualification or a guarantee of model quality and speed. The previous package remains available for rollback.
Developed by 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. See the license and evaluation permission.
Historical project notes and credits describe earlier releases; they do not replace the requirements above.