Browser runtime
N
What can I help with?
Load a GGUF or LittleBit model. Settings are read from its metadata.
NEXUS · Nexus custom inference · General AI
Load a GGUF or LittleBit model. Settings are read from its metadata.
No hardware capacity declared yet. Select your card before loading a large model.
Load a model for automatic memory planning. This is a custom runtime, not llama.cpp parity.
Curated configuration profiles are not a popularity ranking or a guarantee of support. Each file must pass its architecture, tokenizer, tensor and memory checks. ChatGPT service models are not local GGUF files. Experimental Gemma, GPT-oss and Qwen image routes have separate qualification limits; profile coverage does not establish model quality or device speed.
The inherited Nexus WebGPU runtime validates compatible dense graphs at load and includes separately guarded experimental architecture and image routes. Unsupported tensor layouts and templates fail closed. Synthetic numerical and lifecycle tests do not certify full-checkpoint answers, practical memory fit, browser playback or physical-GPU throughput. See the release README and QA receipts for exact validation scope.
Your models. Your preferences. One workspace.
Nexus AI browser trial. Select your GPU memory profile, then load a supported GGUF from your disk. WebGPU is required; no model is bundled. Model files and prompts stay in this browser. Native inference and companion services require the full app.
Load the supported local Qwen2.5-VL-7B text GGUF in the model selector, then its matching local mmproj GGUF here. Custom WebGPU image encoder; numerical fixtures tested. Full 7B checkpoint quality and browser/GPU qualification remain unverified.
Weights stream from the selected local file. Vision allocations share the student/teacher GPU cap. Unload the companion to reclaim its cache.
No image selected.
Image answers use base model weights. Active learned heads and chat memory are not applied.
Select a local Qwen2.5-VL-7B model to begin.
Only the image you explicitly select is used. Camera and expression-cue frames are never attached. The image is reduced to at most 50,176 pixels (224 × 224 for a square), with at most 64 image tokens; fine print may be lost. Transparent pixels are flattened on white.
This panel is temporary: image, question and answer are not saved to conversation history, memory, backups or correction training. Clear removes them; changing models or ending the session also clears the panel.
No network requests are made until you click Check files. Direct downloads use public publisher or clearly identified GGUF-converter repositories. Read each model's licence before downloading.
Runtime memory includes more than the weight file.
Downloads stay on your device. Load the saved GGUF using the model selector. No access tokens are requested or stored.
Gated models, blocked cross-origin requests, or browsers without direct-to-disk support: use the source link to download in your browser, then select the local GGUF. This avoids holding a multi-gigabyte download in JavaScript memory.
Load a GGUF to inspect its compatibility contract.
SmolLM and SmolLM2 have different tokenizer templates, positional encodings and head layouts. Base checkpoints use completion formatting and are not instruction-tuned assistants. The check below reports implementation compatibility; answer quality and hardware throughput require separate evaluation.
No llama.cpp or PyTorch is used for local inference. Model weights remain separate from this HTML. Enter actual VRAM in GPU settings before loading a large model. Changing GPU memory settings requires reloading the model.
| Model family | Local custom engine |
|---|---|
| SmolLM / SmolLM2 | 135M, 360M, 1.7B · base/instruct metadata profiles. Exact tokenizer, tensor types, RoPE and memory checks apply. Profile coverage is not a real-model test pass. |
| Qwen | Dense Qwen2 / Qwen3. MoE and other graph variants rejected. |
| Mistral / Ministral 14B | Conditional on GGUF graph, tokenizer, attention and available memory. Sliding-window variants rejected. |
| GPT-oss 20B / 120B | Experimental MoE / MXFP4 WebGPU backend with Harmony. Bounded numerical tests; full-checkpoint and hardware qualification remain separate. 120B requires upgraded hardware and compatible GGUF files. |
| Gemma 3 | Experimental text graph with local/global attention, Q/K norms and GELU. Other Gemma generations and Gemma image input are not implied. |
| Qwen2.5-VL | Separate experimental multimodal backend and matching local projector required. Check release qualification before use; other Qwen-VL generations are not implied. |
The test sends “Reply with OK” to the configured endpoint. No automatic fallback uploads local conversations.
Speak one phrase, then Nexus answers. The microphone pauses while it thinks and speaks, then resumes automatically after playback. Use Interrupt to speak sooner. This prevents speaker audio from becoming another user message. Speech uses complete sentences and longer clauses for smoother pacing. Automatic browser voice prefers an installed natural voice when available; you can select a different local or online voice below. The local neural voice pack is also available. Starting voice requests microphone permission; it never starts automatically. Offline mode never falls back to online speech.
Edge on Windows: automatic speech prefers an installed local voice. If your saved Read aloud setting is off, enable it once. If Edge blocks sound, use Play reply. Voice test uses the selected speech route.
Natural English speech is generated on this device without an API key after you load a separate compatible Kokoro voice pack. This single-HTML release does not contain voice weights. Choose Load bundled Kokoro and select an existing voice-pack folder, or explicitly use the optional Download button to retrieve missing files from Hugging Face, GitHub and jsDelivr. No conversation text is sent by the download. Strict offline blocks downloads and inference stays local. Browser storage can be cleared or evicted; keep your original pack and exported backups.
Choose Local neural above after installation, then Test voice. English only. Unusual names may be mispronounced. If synthesis fails before playback, Nexus can use an installed local browser voice; it never uploads the text to a fallback service. Spoken responses are AI-generated. Kokoro model/voices: Apache-2.0. HeadTTS G2P and ONNX Runtime: MIT. CMU dictionary: BSD-style. Full notices are included in the app/source documentation.
Speech APIs use the base URL and session key under Models, with /audio/speech and /audio/transcriptions. Browser voices depend on your device; online availability is not guaranteed. Browser recognition may transmit audio to its speech service. API recording is limited to 60 seconds per clip. Spoken responses are AI-generated.
Stores statements such as “I prefer short replies.” No hidden personality diagnosis or background upload. Approved feedback can train the optional output-projection LoRA adapter. Up to 80 facts; recent conversation is bounded. Disabling learning stops new facts. Disable personalization to exclude stored facts from prompts.
No reference notes
Your self-reported mood takes priority. The camera cannot reliably determine your inner emotional state. Voice transcription captures words, not vocal emotion.
No camera frames sent. Camera preview alone remains local.
Live observations require an image-capable API model. Text-only local GGUF models cannot see the camera. Observations expire after 30 seconds and are not saved as memories. This is a Nexus camera interaction, not a LAN person-to-person calling service.
Profile → retrieved sources → feedback examples → optional token-score adapter. Model weights remain separate. No uploads in offline mode.
Camera preview works locally. Speech output needs an installed local voice. Offline dictation needs browser on-device recognition and an already installed language pack. No automatic online fallback.
Documents and learning data are stored in IndexedDB in this browser. Export a backup before moving this HTML or changing browsers. Limits: 100 documents / 10 MB text, 500 feedback items, 3 model adapters. No silent eviction when full.
Choose Not helpful to save a rejected answer, then Add correction to write the complete answer you prefer. Review and explicitly approve the pair before training. Helpful can approve an unpaired positive example. Every correction retains the original answer and its exact saved generation context. Ratings alone never invent a teaching target; praise does not fact-check an answer.
Creates sparse previous-token → next-token scores, conditioned on ordered prompt features. This distinguishes some word-order corrections that token counts miss. The full model file and tokenizer are hashed before binding. Training uses a versioned snapshot and rejects stale results. With five or more distinct prompts, separate held-out preference loss is reported. This is not LoRA, full transformer fine-tuning, DPO or PPO/RLHF. It cannot teach new knowledge reliably; RAG carries your corrections. Quality improvement is unproven.
Runtime controls affect inference, not training. Optional online features are explicitly disabled by the offline lock. Inactive legacy controls have been hidden.
Use Helpful or Correct beneath a reply. Your existing feedback is the only teaching record. Corrections remain available through RAG; the optional adapter trains the loaded model’s output projection using actual transformer activations.
Load an assistant model to prepare local learning.
Review and approve each example, train explicitly, inspect evaluation, then activate explicitly. A thumbs-down alone never creates a target. Corrections train complete preferred/rejected pairs; the CE token limits apply only to unpaired positives.
No candidate evaluated.
This trains output-projection LoRA only; quantized transformer blocks remain frozen. WebGPU computes real hidden states and JavaScript optimizes the head. Approved corrections use full-answer DPO (beta 0.1), comparing preferred and rejected sequence log-probabilities against the immutable raw base model. Unpaired approved positives use cross-entropy. This is not transformer-wide RL, PPO or QLoRA. At least five independent prompt families and two held-out families are needed; preference training also requires two held-out preference families. Complete preferred-answer loss, reference-adjusted pair margin/accuracy, and independent fixed regression checks gate activation. Raw pair accuracy compares sequence log-probabilities and is length-sensitive; preference probability is not calibrated confidence or a correctness score. These limited checks do not prove factual accuracy or general quality. Restored adapters remain disabled until re-evaluated against the exact model and feedback.
Your main loaded model is the student. Load a separate local teacher to write examples, review the corpus, then train a real output-head LoRA. Transformer weights stay frozen. Both models and the full corpus must fit your device and browser storage; 10 million is a token ceiling, not a speed or quality promise.
Load a student in Models first.
Corpus tokens are counted with the student’s actual conversation formatter and include assistant end tokens. Prompt tokens are excluded. Imported passages become approved-answer candidates. Generation is local and stops at the target. No generated example is approved automatically.
No corpus yet.
Training streams at most eight actual hidden-state/logit pairs at a time. Five or more exchanges reserve every fifth exchange for evaluation. Evaluation samples up to16 tokens; smaller corpora provide a training-fit check only. Better token loss does not establish better reasoning or truthful answers.
Activation needs at least five genuinely different scenario families and two held-out families. Numbered variations and duplicates stay together. A single broad teaching topic remains one family: add distinct substantive scenarios, rather than renaming duplicate examples. Full-answer and fixed regression checks measure limited token loss, not general quality.
This is the NXBB0001 Big Brother checkpoint: LoRA weights, Adam optimizer, RNG, progress, model fingerprints and corpus hash. It is not a v11 NQAT0001 recovery checkpoint. For another browser, export both corpus JSONL and .nxbb (older NXBB0001 .nqat exports remain readable), reload the exact student model, import and approve the corpus, then import the snapshot. Teacher weights are only required when generating more examples. Base-model files are never included.
Ready. Start by loading a student and teacher.
Give your model a repeatable check-up. The first completed test is saved as its fixed baseline. Repeat the same test after learning to compare results fairly.
This text stays separate from automatic training, RAG and Big Brother corpus creation. Whether a model saw similar text during its earlier training is unknown. Use your own unseen text for a more useful check.
Raw next-token probabilities, with no sampling, chat template, RAG, personality prompt, or output-score preferences. Every target token is counted once across overlapping windows. Changing the text, tokenizer, context, stride or adapter creates a separate baseline. Maximum: 65,536 text tokens.
Baselines are fixed records in this browser, tied to the complete model fingerprint. Export them before moving this file or clearing browser data. Imports never replace an existing first baseline. Imported measurements cannot be independently verified.
Choose a password to keep everyday users out of settings. This is a local interface lock, not encryption or operating-system security. Someone with browser developer tools or access to this HTML and browser storage can bypass it. No password is set by default.
Checking the settings lock…
A new browser session always starts with settings locked when a password exists. Chat, the microphone, and the camera remain available. Clearing this browser's site data removes its settings lock and may also delete saved learning data; export the data you need before doing that.
Required Vision model, privacy and accuracy notices are appended separately and cannot be removed by editing this text. A model or notice change requires fresh review. This startup acknowledgement does not grant camera, biometric, storage, subject or sharing consent. The original disclaimer is retained so you can restore it. When startup display is off, Nexus skips it only after this browser has explicitly accepted this exact revision. Editing the text or turning startup display back on requires a new acceptance at the next startup. Declining always stops Nexus.
General AI only. The Nexus parser, packed LittleBit kernels, tokenizer and WebGPU graph are included in this file. Online provider compatibility is separate from local model support. LLM weights stay separate. Optional local Vision uses pinned YuNet, SFace, Progressive Teacher FER and YOLOX model sidecars. Facial identity, calibrated expression cues, common objects and cats/dogs are independently controlled under Settings → Vision. Keep the complete assets folder beside this HTML and use localhost or HTTPS.
Chat context keeps the configured number of recent successful turns (default 8, maximum 24); export and local conversation recovery keep up to 100 recent turns, and the screen keeps 100 messages. Start a new conversation to clear its transcript. Persistent profile, explicit preference learning, reference-note retrieval, chat export, voice controls, camera preview, online live observations and stop/recovery are integrated. The original Python app’s distributed workers, LAN messenger, OS loopback routing, voiceprint targeting and scheduled OS tasks are not ported. Learning trains a frozen-base output-projection LoRA adapter from approved answers. Big Brother uses a separate teacher model to build a reviewed corpus and train the same adapter type. Workspace backups include settings, conversation, feedback, LoRA and perplexity records. Large Big Brother corpora and recovery checkpoints have separate exports; credentials and model weights are excluded.
Declining closes script-opened windows where allowed. Otherwise Nexus locks and stops; browsers do not let a file close an arbitrary user-opened tab.
Regression checks cover tokenizer IDs and supported templates, settings, retrieval, storage recovery, cancellation, simulated camera/microphone and voice turn-taking. Provider responses and model generation are mocked for integration checks; real-model answer quality, Radeon performance, microphone accuracy and live provider compatibility require testing on your device.
Learning checks cover versioned training snapshots, feedback deletion, full-file binding, adapter activation and sequence-sensitive updates. Controlled numeric tests do not establish better full-model answers.
Learning uses actual transformer hidden states to train an output-projection LoRA adapter. Transformer blocks and the source GGUF stay frozen. Adapters bind to the full model hash and require evaluation before activation. This is supervised adaptation from approved examples, not full-model RLHF.
Background research: LoRA (Hu et al.) and DPO (Rafailov et al.).
Nexus / Nexus — user-supplied CAF AI Assistant and Nexus Web LLM source. Model weights remain subject to their own licenses. SmolLM2: Hugging Face. Mistral: Mistral AI. Qwen: Qwen team. LittleBit-inspired factorization remains experimental.
Changes save in this browser. API keys are session-only.
Approval allows this exact record to be used when you explicitly train a candidate. Approval does not train or activate an adapter. Corrections compare your preferred answer with the rejected answer under the same saved generation context.