The whole platform
A private AI studio that replaces a stack of cloud subscriptions: chat, image, audio, video, 3D, agents, training, and a full model catalog — behind one login, on a private GPU fleet, where your prompts, documents and fine-tunes are never sent to a third-party model API. This page is the complete tour. Where something isn't built yet, it says so.
Every cloud AI tool is the same bargain: convenience now, your data and your dependence forever. Nexus is the other choice — the same frontier capabilities, with sovereignty over the two things that are actually yours: your data and your work.
Inference, training and storage all run inside your own cluster. No prompt, document or weight is sent to a third-party API. For law, medicine and finance, "did anyone outside the firm see that?" has one answer: no.
Chat, images, voice, music, video, 3D, agents and fine-tuning in one workspace on one meter — instead of a separate monthly bill, login and data-sharing agreement for each.
Pull any open model, fine-tune it on your corpus, fuse and quantize it — your fine-tunes, prompts and outputs are yours to keep and export. You are not renting access to a model that can be deprecated or repriced out from under you.
The Sovereign gate mediates every write, delete, command and network call through a tiered permission model, with consent prompts and an append-only audit trail. Nothing destructive happens silently.
Each workspace runs in its own sandbox with no path to the control plane or to anyone else's data. Tenants get a full dev environment; the backend is a no-client zone.
Start on the shared fleet, bring your own GPU on the desktop app, or stand the whole platform up on your own hardware — on-prem and air-gapped. The software is the same; only where it runs, and who owns it, changes.
A full private workspace, not a single endpoint. Each hub is toggleable per tenant from the admin console, and every one ships with a Guided and a Technical walkthrough. Intake, Training and model downloads start at the Studio tier.
Models ready, jobs running, GPUs busy, what finished today, quick launch.
An agent plans and runs a goal step by step; Companions remember; every send or delete waits for your yes.
Chat with vision and vault-backed memory — models sized to your plan, on a fleet benchmarked to 123B-class.
SDXL / FLUX-class generation with ControlNet, batch and post-processing.
Text-to-speech, transcription, music and short video generation.
Model, sculpt, UV, rig and slice for print — a full 3D pipeline in the browser.
Pull models, repos and datasets; index your document vaults for retrieval.
LoRA / QLoRA fine-tunes, MergeKit fusion and evals — your weights stay yours.
See the shared models on your plan and bind plugins; from Studio up, download from Civitai, Tensor.Art and Hugging Face into private storage.
Every output, asset, file and reference in one versioned, searchable grid.
Jobs, monitors, GPU stack, disk saver and network — with honest live progress.
Theme, Advanced Mode, and Guided or Technical tutorials for every hub.
Operators get two more surfaces — Ops and Admin — for GPU and job state, tenants and keys, monitors, the audit trail and approvals. Spatial / XR surfaces are in active development.
The Agents hub builds first-party agents that plan, use tools, and run jobs on your behalf — each one a persona, a model, a tier, and a precise set of capabilities you switch on.
Agents draw from a catalog of 76 tools, 59 MCP connectors (GitHub, databases, cloud, browsers, docs, search, and more), and a plugin and skill library. Nothing is on until you turn it on, per agent. When a connector needs a key, you're told exactly where to get it — and the secret only ever lives in your backend, never in the browser.
Every capability carries a risk tier. Offensive and defensive security tooling is operator-only. Writes, deletes and commands pass the Sovereign gate and can require a consent tap. A tenant works in a full isolated dev environment with the normal dev toolset; the control plane stays a no-client zone.
Give an agent a persona from a preset or your own prompt. Import SillyTavern-style character cards, or create and generate new ones from a brief — then turn any card into a working agent in one click.
Companions keep long-term, vault-backed memory. Agents run on a schedule or on demand, queue jobs, and drop results in a task inbox — with every send or delete held for your approval first.
The generative tools people pay for one at a time, in one place, on private GPUs we own — so the work and the source material both stay private.
SDXL / FLUX-class generation, ControlNet guidance, inpainting and post, batch pipelines, and an asset library that remembers every recipe.
Text-to-speech and voice, transcription, and music generation — drafted, versioned and kept in your library.
Short video generation from text and stills, with the same job tracking and honest progress as everything else.
A browser 3D pipeline — model, sculpt, UV, rig, and slice for print — not a viewer, a workbench.
Download from Civitai, Tensor.Art and Hugging Face straight into your private storage, with a search that shows what you already have.
LoRA / QLoRA training on your own data. Fuse models with MergeKit, quantize, and evaluate — the output weights land on your disk.
Every model — shared or your own — on a single OpenAI-shaped endpoint and a single usage meter, with plugins bound per model.
The full desktop-class studio in the browser — twelve hubs, dockable panes, ⌘K palette, session recovery.
Windows, Linux and macOS. Runs a local backend, and can bring your own GPU for local compute. Studio tier and up.
A phone client for your Nexus — sign in with your key (sealed in the device key store, entitlement verified on-device) and work from the hubs that are wired for mobile, with more arriving. Solo+.
An OpenAI-shaped, streamed API for your own tools, plus a full isolated Devbox terminal — your sandbox, no path to ours.
iOS isn't built yet — no dishonest "coming soon" date. Spatial / XR surfaces are in active development.
Inference, training and storage stay inside your cluster. Nothing is forwarded to a third-party model API.
One tiered permission gate on every write, delete, command and fetch, with consent prompts and a redline for the most dangerous actions.
An append-only audit log of every privileged action and notice — built for the questions auditors actually ask.
Each workspace is sandboxed with its own storage and no access to the control plane or to another tenant's data.
Pause or cut compute instantly from the operator console when you need to.
For law, medicine and finance: run the whole platform on your own hardware, disconnected from the internet entirely.
Every number below is a real device benchmark, run in-cluster from the control plane. No projections, no marketing math. Where something isn't built yet, we say so.
Benchmarked 2026-09-24 · FP16 8192³ matmul · serving on Qwen2.5-7B (Q4), GPU-resident
| Tier | Cards | Arch | FP16 | VRAM |
|---|---|---|---|---|
| NVIDIA — workhorse | 8× RTX 5060 Ti | Blackwell (sm_120) | ~50 TFLOP/s ea | 16 GB ea |
| NVIDIA — Ada | 1× RTX 4060 Ti | Ada (sm_89) | 45.8 TFLOP/s | 16 GB |
| AMD — modern | 1× RX 9060 XT | RDNA4 (gfx1200) | 40.4 TFLOP/s | 16 GB |
| AMD — HBM2 | 2× Radeon Pro VII | gfx906 | 15.5 + 16.6 TFLOP/s | 16 GB ea |
| Verified total · 12 GPUs | ~520 TFLOP/s | 201 GB | ||
Real token generation on a 7B (Q4) model, GPU-resident.
Serving pool = 7 Blackwell + 1 Ada + 1 RDNA4 for fast 7B. The RX 9060 XT and both HBM2 Radeon Pro VII cards are cluster-joinable as GPU peers via Vulkan (RADV) — folding the full fleet into one large-model pool.
Every worker clears the ~30B (4-bit) bar on one card; the NVIDIA tier does it eight times over, in parallel. 70B-class has been benchmarked across a cross-GPU llama.cpp RPC pool (5 GPUs, measured ~7 tok/s) over the 2.5 GbE fabric — modest speed, but real. Larger models are in active bring-up.
The nodes are separate machines joined by a 2.5 GbE fabric — ideal for running many models side by side and serving many users in parallel: 12 independent GPU endpoints, a throughput farm. Splitting one very large model across nodes is bottlenecked by the network, not the silicon; we built exactly that path — a llama.cpp RPC pool (CUDA + Vulkan) that has run 70B- and 123B-class models (benchmarked) at modest, honest tok/s. Larger models are served on request, not by default.
Start free on the shared fleet, bring your own GPU, or run the whole platform on-prem. The same studio, wherever it runs.