Repository navigation
Welcome to NEXUS AI Studio
Version 25
English · Français · 日本語 · 한국어
Open the browser trial · Download the full application · README
A local AI workspace for chat with supported language models through its custom WebGPU or Nexus.cpp inference engine.
Open the trial, choose your GPU profile and memory capacity in Models & settings, then select a local GGUF model and load it for chat. No model is bundled. The custom WebGPU engine checks the model architecture, tensor graph, quantization and tokenizer. A working WebGPU adapter and enough memory are required; selecting a GPU profile does not prove compatibility or measure free VRAM. RX 7900 XT and XTX presets represent nominal 20 GB and 24 GB capacities.
The single HTML file embeds the interface, inference code and tokenizer/template library for online evaluation without Python, installation or a build step. Model import and loading take additional time. The trial does not upload model data, use a local server or fall back to CPU inference. Native Nexus.cpp, online services, voice, camera and teacher features are not included.
The full local application integrates model checks, controllable conversation state and inspectable diagnostics across its separate browser and native engines. Native protocol 3 provides Stop, New Chat, prefix caching and sampler controls; native executables must be rebuilt for those features. Native batch remains one token per step; native Flash Attention and reduced-precision KV are not implemented.
Download the complete package and keep its numbered data files, manifest, installer and supplied layout. Run Install_Nexus_AI.py with CPython 3.10+; recorded Linux checks used 3.12.14. Nexus AI does not use a PyTorch backend. The full application needs its local service; Nexus.cpp needs a compatible executable and the appropriate compiler/HIP/driver setup for Radeon acceleration. The optional desktop window uses pywebview>=5,<7.
The browser packaging has passed scripted no-network boot checks. These are not full-model GPU quality or performance qualification. A finite/nonconstant-logit load check is only a basic readiness check. Windows/Radeon model performance and universal model compatibility are not claimed. Browser revision: v25-browser-r1. The native Version 25 package, Candidate25 paths and hashes remain unchanged; Candidate24 remains available for rollback.
Personal use and internal commercial testing and evaluation are permitted for Nexus-owned material, including the new browser packaging. Distribution, modification, production deployment and other commercial use require express written consent. Third-party licenses and previously granted rights remain unchanged.
Rich Dunbar / Nexus Emerging Technology. License · Evaluation permission (English).
Ouvrir la version navigateur · Télécharger l’application complète · README
Un espace IA local pour dialoguer avec les modèles pris en charge via le moteur personnalisé WebGPU ou Nexus.cpp.
Ouvrez l’application, choisissez le profil GPU et la capacité mémoire dans Models & settings, puis sélectionnez un modèle GGUF local et chargez-le pour dialoguer. Aucun modèle n’est inclus. Le moteur WebGPU vérifie l’architecture, le graphe des tenseurs, la quantification et le tokenizer. Un adaptateur WebGPU fonctionnel et une mémoire suffisante sont nécessaires. Un profil GPU ne prouve pas la compatibilité et ne mesure pas la VRAM libre. Les profils RX 7900 XT et XTX indiquent des capacités nominales de 20 Go et 24 Go.
Le fichier HTML unique intègre l’interface, le code d’inférence et la bibliothèque de tokenisation et de modèles de conversation pour permettre une évaluation en ligne sans Python, installation ni compilation. L’importation et le chargement du modèle prennent du temps supplémentaire. La version navigateur ne téléverse pas les données du modèle, n’utilise pas de serveur local et ne propose aucun repli vers le CPU. Nexus.cpp natif, services en ligne, voix, caméra et fonctions d’enseignant ne sont pas inclus.
L’application complète associe vérification des modèles, contrôle de la conversation et diagnostics consultables dans ses moteurs navigateur et natif distincts. Le protocole natif 3 propose Stop, New Chat, cache de préfixe et réglages d’échantillonnage ; les exécutables natifs doivent être recompilés pour ces fonctions. Le lot natif reste limité à un token par étape ; Flash Attention natif et cache KV à précision réduite ne sont pas implémentés.
Téléchargez le paquet complet en conservant les fichiers de données numérotés, le manifeste, l’installateur et l’arborescence fournie. Exécutez Install_Nexus_AI.py avec CPython 3.10+ ; les contrôles Linux enregistrés utilisent 3.12.14. Nexus AI n’utilise pas de backend PyTorch. L’application complète exige son service local ; Nexus.cpp exige un exécutable compatible et une configuration compilateur/HIP/pilote adaptée à l’accélération Radeon. La fenêtre de bureau facultative utilise pywebview>=5,<7.
La version navigateur a passé des contrôles scriptés de démarrage sans réseau. Ils ne qualifient pas la qualité ni les performances GPU d’un modèle complet. Un contrôle de logits finis et non constants n’est qu’un test élémentaire de disponibilité. Aucune performance Windows/Radeon ni compatibilité universelle n’est revendiquée. Révision navigateur : v25-browser-r1. Le paquet natif version 25, les chemins Candidate25 et les empreintes restent inchangés ; Candidate24 reste disponible pour revenir en arrière.
L’usage personnel et les tests et évaluations commerciaux internes sont autorisés pour les éléments appartenant à Nexus, y compris cette version navigateur. La distribution, la modification, le déploiement en production et les autres usages commerciaux nécessitent un consentement écrit exprès. Les licences tierces et les droits déjà accordés restent inchangés.
Rich Dunbar / Nexus Emerging Technology. License · Evaluation permission (English).
ブラウザー試用版を開く · 完全版アプリをダウンロード · README
対応モデルを独自の WebGPU または Nexus.cpp 推論エンジンで使用する、ローカル AI チャット環境です。
試用版を開き、Models & settings で GPU プロファイルとメモリー容量を設定した後、ローカルの GGUF モデルを選択して読み込むとチャットを試せます。モデルは同梱していません。WebGPU エンジンがアーキテクチャ、テンソル構造、量子化、トークナイザーを検証します。動作する WebGPU アダプターと十分なメモリーが必要です。GPU プロファイルの選択は互換性の証明や空き VRAM の測定ではありません。RX 7900 XT と XTX の設定値は、公称容量の 20 GB と 24 GB です。
単一 HTML に画面、推論コード、トークナイザーと会話テンプレートのライブラリーをまとめ、Python、インストール、ビルドなしでオンライン評価を始められるようにしています。モデルの選択と読み込みには別途時間がかかります。モデルデータのアップロード、ローカルサーバーの使用、CPU 推論への代替処理は行いません。ネイティブ Nexus.cpp、オンラインサービス、音声、カメラ、教師機能は含みません。
完全版は、独立したブラウザーエンジンとネイティブエンジンに、モデルの検証、会話状態の制御、確認可能な診断情報を統合します。ネイティブプロトコル 3 は Stop、New Chat、プレフィックスキャッシュ、サンプリング設定を提供します。これらの機能にはネイティブ実行ファイルの再ビルドが必要です。ネイティブのバッチは 1 ステップ 1 トークンで、ネイティブ Flash Attention と低精度 KV キャッシュは未実装です。
完全版の全ファイルをダウンロードし、番号付きデータファイル、マニフェスト、インストーラー、フォルダー構成を保持してください。Install_Nexus_AI.py には CPython 3.10 以上が必要です。記録済みの Linux 検証では 3.12.14 を使用しました。Nexus AI は PyTorch バックエンドを使用しません。完全版にはローカルサービスが必要です。Nexus.cpp の Radeon アクセラレーションには対応する実行ファイルとコンパイラー、HIP、ドライバー構成が必要です。任意のデスクトップウィンドウは pywebview>=5,<7 を使用します。
ブラウザー版はネットワーク通信なしの起動スクリプト検証を通過しています。これはモデル全体の GPU 品質・性能の認定ではありません。有限かつ一定でないロジットの確認は、基本的な準備状態の検査にすぎません。Windows/Radeon の性能や全モデルへの対応は保証していません。ブラウザー改訂番号は v25-browser-r1 です。ネイティブ版 25 のパッケージ、Candidate25 のパスとハッシュは変更せず、Candidate24 もロールバック用に残します。
新しいブラウザー版を含む Nexus 所有部分は、個人利用および社内の商用テスト・評価に利用できます。配布、改変、本番導入、その他の商用利用には明示的な書面による同意が必要です。第三者のライセンスと既に付与された権利は変更されません。
Rich Dunbar / Nexus Emerging Technology. License · Evaluation permission (English).
브라우저 체험판 열기 · 전체 애플리케이션 다운로드 · README
맞춤형 WebGPU 또는 Nexus.cpp 추론 엔진으로 지원 모델을 사용하는 로컬 AI 채팅 작업 공간입니다.
체험판을 열고 Models & settings에서 GPU 프로필과 메모리 용량을 설정한 다음 로컬 GGUF 모델을 선택하여 불러오면 채팅을 시험할 수 있습니다. 모델은 포함되어 있지 않습니다. WebGPU 엔진이 아키텍처, 텐서 구조, 양자화 및 토크나이저를 검사합니다. 작동하는 WebGPU 어댑터와 충분한 메모리가 필요합니다. GPU 프로필 선택은 호환성을 입증하거나 남은 VRAM을 측정하지 않습니다. RX 7900 XT와 XTX 설정은 공칭 용량인 20 GB와 24 GB입니다.
단일 HTML에 인터페이스, 추론 코드, 토크나이저 및 대화 템플릿 라이브러리를 포함하여 Python, 설치 또는 빌드 없이 온라인 평가를 시작할 수 있습니다. 모델 선택과 로딩에는 추가 시간이 필요합니다. 체험판은 모델 데이터를 업로드하거나 로컬 서버를 사용하지 않으며 CPU 추론으로 전환하지 않습니다. 네이티브 Nexus.cpp, 온라인 서비스, 음성, 카메라 및 교사 기능은 포함하지 않습니다.
전체 앱은 별도의 브라우저 및 네이티브 엔진에 모델 검사, 대화 상태 제어와 확인 가능한 진단 정보를 통합합니다. 네이티브 프로토콜 3은 Stop, New Chat, 접두사 캐시 및 샘플링 설정을 제공합니다. 이 기능을 사용하려면 네이티브 실행 파일을 다시 빌드해야 합니다. 네이티브 배치는 단계당 토큰 하나이며 네이티브 Flash Attention과 저정밀 KV 캐시는 구현되지 않았습니다.
전체 패키지를 다운로드하고 번호가 붙은 데이터 파일, 매니페스트, 설치 프로그램과 폴더 구성을 유지하세요. Install_Nexus_AI.py는 CPython 3.10 이상이 필요하며 기록된 Linux 검사는 3.12.14를 사용했습니다. Nexus AI는 PyTorch 백엔드를 사용하지 않습니다. 전체 앱은 로컬 서비스가 필요합니다. Nexus.cpp의 Radeon 가속에는 호환 실행 파일과 적절한 컴파일러/HIP/드라이버 구성이 필요합니다. 선택적 데스크톱 창은 pywebview>=5,<7을 사용합니다.
브라우저 패키징은 네트워크 통신 없이 시작하는 스크립트 검사를 통과했습니다. 이는 전체 모델의 GPU 품질이나 성능 검증이 아닙니다. 유한하고 일정하지 않은 로짓 검사는 기본 준비 상태 확인에 불과합니다. Windows/Radeon 성능이나 모든 모델과의 호환성을 보장하지 않습니다. 브라우저 개정 번호는 v25-browser-r1입니다. 네이티브 버전 25 패키지, Candidate25 경로와 해시는 변경하지 않았으며 Candidate24도 롤백용으로 유지합니다.
새 브라우저 패키징을 포함한 Nexus 소유 자료는 개인적 이용과 내부 상업적 테스트 및 평가에 사용할 수 있습니다. 배포, 수정, 운영 환경 배포와 그 밖의 상업적 이용에는 명시적인 서면 동의가 필요합니다. 제3자 라이선스와 이미 부여된 권리는 변경되지 않습니다.
Rich Dunbar / Nexus Emerging Technology. License · Evaluation permission (English).
The original text below is project history, not current installation, runtime or licensing guidance. Use the current Version 25 section above and its linked documentation.
Le texte original ci-dessous est l’historique du projet, pas un guide actuel d’installation, d’exécution ou de licence. Utilisez la section actuelle de la version 25 ci-dessus et sa documentation liée.
以下の原文はプロジェクトの履歴であり、現在のインストール、実行環境、ライセンスの案内ではありません。上記のバージョン 25 の説明とリンク先の資料を使用してください。
아래 원문은 프로젝트 이력이며 현재 설치, 실행 환경 또는 라이선스 안내가 아닙니다. 위의 버전 25 설명과 연결된 문서를 사용하세요.
Developed by Rich Dunbar with AI-assisted coding and hundreds of build–measure–learn iterations.
NEXUS AI Studio brings custom browser-based inference, optional voice and expression modules, retrieval-based memory, controlled adapter learning and perplexity evaluation into one application.
The project documentation is available in:
- English
- Français — French
- 한국어 — Korean
- 日本語 — Japanese
Select your preferred language from the wiki’s page list. Each version covers the development history, architecture, implemented model support, learning features, testing evidence, limitations and research credits.
La documentation française présente le développement, l’architecture, les fonctionnalités, les limites et les références scientifiques.
한국어 문서에서는 개발 과정, 아키텍처, 기능, 한계와 연구 출처를 확인할 수 있습니다.
日本語版では、開発の経緯、構成、機能、制約、研究の出典をご覧いただけます。
I wanted a customizable assistant that could run on my own hardware and fit my workflow.
Development began with LM Studio, progressed through a custom graphical interface and evolved into my own browser-based inference implementation. Moving from executable packaging toward HTML made interface changes and distribution easier for my development process.
The current local inference path uses custom JavaScript, WebGPU and WGSL code without requiring LM Studio or llama.cpp. Optional online services are separate.
This project grew through repeated implementation, testing, screenshots, diagnostic information and end-user feedback.
flowchart LR
A["Build"] --> B["Test and measure"]
B --> C["Learn from evidence"]
C --> D["Refine"]
D --> A
E["User feedback"] --> C
F["Screenshots and diagnostics"] --> B
style A fill:#123456,stroke:#38bdf8,color:#ffffff
style B fill:#164e63,stroke:#22d3ee,color:#ffffff
style C fill:#312e81,stroke:#a5b4fc,color:#ffffff
style D fill:#14532d,stroke:#4ade80,color:#ffffff
A working minimum viable product provides something real to test. Continued feedback and regression checks guide improvements in reliability, clarity and capability.
The main interface and logic share an HTML entry point. The complete offline package also contains separate neural-model and runtime assets.
| Area | What the inspected build implements |
|---|---|
| Local inference | Custom WebGPU execution for specifically supported model profiles |
| Memory and retrieval | Context supplied without necessarily changing model weights |
| Speech output | Included English Kokoro voice pack for local synthesis |
| Speech input | Browser-dependent recognition; strict offline support has additional requirements |
| Expression-aware interaction | Optional, uncertain expression cues for cautious tone adaptation |
| Preference learning | Approved text examples and output-head LoRA-style adaptation |
| Evaluation | Held-out promotion checks and protocol-aware perplexity scoring |
The app can adapt an output-head adapter using approved answers and preference pairs. The base model and transformer blocks remain frozen.
Training, evaluation and activation controls help keep changes deliberate. Optional end-of-session learning can run when eligible approved data and a verified supported model are ready. New adapters must pass evaluation gates before promotion.
Ordinary conversation does not automatically retrain the whole model, add parameters or guarantee better performance.
The camera module estimates facial-expression categories. It does not establish a person's inner feelings, identity, intentions or mental health.
An optional, default-off local-chat feature can use a recent estimate as a cautious prompt cue. The user's own words and self-reported mood take precedence. Camera-influenced contexts are excluded from automatic adapter-training capture.
The included Kokoro route provides local English speech output after its assets are imported.
Speech recognition is a separate feature. Strict offline recognition requires compatible on-device browser support and an installed language pack.
Offline output does not automatically mean offline input. Online API and browser-service options must be considered separately.
Implemented support is specific to model architecture, tensor format, tokenizer, template and memory requirements. A model name in the interface does not establish full compatibility.
The detailed pages distinguish:
- Implemented code paths
- Synthetic and numerical software checks
- Limited historical real-model checks
- My reported Radeon RX 7900 XT, Windows 11 and Microsoft Edge workbench experience
- Features and hardware combinations still needing validation
Perplexity can help compare compatible checkpoints under identical conditions. It is not a complete measure of intelligence, factual accuracy or reasoning.
The detailed documentation credits LM Studio, model creators, LoRA, direct preference optimization, retrieval-augmented generation, affective-computing research and the vision and voice dependencies.
Academic citations explain the underlying methods. They do not certify NEXUS AI's implementation or performance.
The inspected application package declares UNLICENSED, and some expression-model provenance questions remain unresolved. Consult the repository's actual licensing information before assuming redistribution or commercial-use permissions.
Useful feedback includes:
- Application version and browser.
- Hardware, driver and backend details.
- Exact model and quantization.
- Reproduction steps.
- Expected and observed behavior.
- Relevant screenshots and diagnostic logs.
Remove private prompts, personal information and secrets before sharing.
Build. Measure. Learn. Verify. Refine.
Rich Dunbar · NEXUS Emerging Technology