Self-serve inference for individual builders, and dedicated on-premise AI for law, medical, and financial practices. Your data never touches a third party; your models run on hardware you can audit. HIPAA-, GLBA-, and attorney-client-privilege ready.
From $39/mo · live GPU-minute calculator
BAA/GLBA · dedicated GPUs · audit packs
185 GB VRAM (measured), air-gapped WireGuard mesh, encrypted per-tenant NAS
"Did anyone outside the firm see the contents of this prompt?"
Inference runs on hardware in your facility — or ours, behind a firewall you control. No prompts, no completions, no telemetry leaves the perimeter.
Every privileged action is logged with a per-tenant trail (7-year retention for legal/SEC, 6-year HIPAA). Hand the JSONL or signed PDF directly to your auditor.
No GPU is ever shared across firms in the privacy tiers. Encrypted scratch volumes wipe at job end. Cross-tenant access denied at three independent enforcement layers.
We run Llama, Qwen, Mistral, MedLlama, BioMistral, DeepSeek, and your fine-tuned variants. We do not train on your data. We do not retain your data.
QLoRA on a 7B–13B model in 1–3 hours on your own briefs, charts, contracts. The trained adapter stays on your tenant — never published, never reused. Larger-model fine-tuning is on the roadmap.
Three independent off-paths: soft suspend, hard revoke, polkit-gated tier escalation with typed authorization for system-level changes. You stay in control.
A private, OpenAI-shaped inference API on hardware you can audit. Every number below is a measured in-cluster benchmark — not a projection.
dense FP16 · 11 GPUs
GPU memory, fleet-wide
9-GPU serving pool
independent GPU endpoints
POST /llm/chat returns OpenAI-shaped chunks token-by-token over Server-Sent Events — read incrementally or buffer to the final message.
GET /llm/catalog lists the models your workspace can reach. Serving Qwen2.5-7B today; 30B-class (4-bit) fits on a single card.
Two HBM2 cards are held for compute and embeddings alongside the token-serving pool — vector workloads don't steal from your chat throughput.
Fine-tuning runs only on your instruction and stays inside your tenant. We never train shared models on your data.
Usage is metered in GPU-minutes across every model. GET /billing/<tid>/summary returns your live usage any time.
Per-tenant isolation, US-hosted, no prompt text in the billing ledger. Firms get a monthly PDF audit pack for them and, if invited, their auditor.
List your models
curl https://api.nexusaisolutions.io/llm/catalog \
-H "Authorization: Bearer sov-YOUR_KEY"Check your usage
curl https://api.nexusaisolutions.io/billing/<tid>/summary \
-H "Authorization: Bearer sov-YOUR_KEY"A Nexus workspace is a full private studio, not a single endpoint. These are the apps provisioned with every customer tenant — each one toggleable per seat from your admin console.
Studio · llm
OpenAI-shaped chat & completions against your own models, streamed over SSE. This is the benchmarked serving path — Qwen2.5-7B today, 30B-class on a single card.
Studio · image
Text-to-image and image-to-image on dedicated GPU workers, entirely inside your tenant.
Studio · a/v
Generate and edit audio and video clips on their own worker pools — nothing leaves the cluster.
Studio · 3d
3D asset generation and scene tooling for pipelines that need geometry, not just text.
Studio · rag
Vector search over your own documents, wired straight into chat — your corpus stays in-tenant.
Audio-gen
Music and sound-effect synthesis pipelines, queued alongside your other generations.
Vision
Image understanding with vision-capable models — describe, extract, and reason over images.
Agents · companion
A persistent assistant with its own memory and task templates, scoped to your workspace.
Agents · forge
An in-browser 3D editor — modeling, sculpt, UV and print slicing, running on the shared runtime.
Core · library
Every generation saved, versioned and searchable — a private asset store per tenant.
Core · jobs
One queue for every long-running generation and training run, with live status.
Spatial / XR surfaces are in active development and aren't counted above. See the measured capabilities →
Self-serve, no sales call. The API is OpenAI-shaped, so Cursor, Continue, Open WebUI and your own RAG harnesses work against your private endpoint. Regulated firm? Enterprise onboarding adds a signed BAA/GLBA and dedicated hardware.
Free trial, no card. You land in your tenant portal with a one-time primary key.
Your primary key (sov-…) is issued instantly. Mint scoped keys from the portal.
One base-URL change. Your key as a Bearer token against /llm/chat.
Upgrade tiers self-serve. Firms add a monthly compliance evidence pack and a signed BAA.
# Stream a chat completion — OpenAI-shaped, over SSE
curl -N https://api.nexusaisolutions.io/llm/chat \
-H "Authorization: Bearer sov-YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{
"backend": "local-direct",
"model": "qwen2.5-7b-instruct",
"messages": [{"role":"user","content":"Summarize this deposition."}]
}'
# Your existing tools, your private model, your data stays in-tenant.Start self-serve in minutes, or run dedicated for regulated work. Every plan runs on private, US-hosted GPUs — your data never leaves the cluster.
Self-serve · start in minutes
Indie developers & solo builders
Small teams & studios
Agencies & growing businesses
All self-serve plans include a free 7-day trial (no card required). Usage past your monthly bucket is $0.02 / GPU-minute, or auto-upgrade to the next tier — your choice. Extra seats $15/mo.
Regulated & dedicated
Dedicated capacity, signed agreements, and on-prem or air-gapped deployment for firms that answer to regulators.
Regulated firms · dedicated hosted capacity
Own the box · custom-quoted
Custom fine-tuning projects: $5,000 – $25,000 one-time, typically 2–6 weeks. On-prem hardware deployment available — we install and operate the cluster at your facility, you own the box.
Our infrastructure was designed compliance-first. The technical posture lines up directly with the regulatory requirements you operate under.
| Audit question | How Nexus answers |
|---|---|
| Does our data leave the premises? | No. Inference, fine-tuning, storage all on-cluster. WireGuard mesh has no exit route to public internet by default. |
| Can your engineers read our prompts? | No. AppArmor enforce on the daemon denies access to per-tenant directories. Per-tenant audit log captures every privileged access attempt. |
| How do you prevent cross-tenant leakage? | GPU scheduler refuses cross-tenant co-location by default. Encrypted ephemeral scratch per inference job, wiped at job end. |
| What if something goes wrong — can you stop it? | Three independent kill paths: soft suspend, hard revoke, polkit-gated tier-3 / tier-4 escalation with typed authorization for system-level changes. |
| Do you sign a Business Associate Agreement? | Yes — standard healthcare-IT BAA on file. GLBA service-provider agreement for financial firms. Custom terms for SEC/FINRA-regulated entities. |
| What about audit trails? | Every privileged action lands in audit.jsonl. 7-year SEC/FINRA retention, 6-year HIPAA, configurable per tenant. Signed PDF evidence pack on demand. |
| Are you SOC 2 compliant? | Type II audit in progress. Technical controls already meet CC6.1, CC6.6, CC6.7 (logical access, system monitoring, software acquisition). |
Our primary cluster is in a secured US facility. For on-prem clients, the hardware can be installed at your facility — we operate it remotely with your physical access controls. We can also do fully air-gapped deployments where the cluster has no internet connection at all.
Llama 3.2 (8B, 70B), Qwen 2.5 (7B, 32B, 72B), Mistral, DeepSeek V3, MedLlama, BioMistral — basically any open-weight HuggingFace model that fits the VRAM. We can pre-download and serve a model whitelist specific to your tenant, or you can bring your own fine-tuned weights and we'll deploy them.
For a 7B-class model we measure ~85 tokens/sec single-stream and ~130 tokens/sec at 4-way concurrency per GPU, prefill above 1,200 tokens/sec. Larger models (up to ~30B at 4-bit) run slower per token; 70B is on the roadmap. On a local network, latency is purely model speed — no public-internet round trip.
Yes. Dedicated/Enterprise plans can include fine-tuning; self-serve plans can purchase fine-tunes as one-off projects. Your training data lives in your encrypted tenant directory, the resulting adapter never leaves it, and we never train on your data without explicit written authorization for that specific purpose.
Yes — required for dedicated / Enterprise plans when handling PHI. Our BAA template uses standard HHS language. We can also accept a customized BAA from your firm's counsel.
For on-prem deployments, you own the hardware and have full root access — your cluster keeps running. For hosted plans, our terms include a 30-day data egress + transition period plus a code-escrow option for the orchestration software so you can migrate to your own infrastructure.
A laptop caps at ~8B and throttles under sustained load. Nexus is 185 GB of measured VRAM across 11 GPUs — enough for 30B-class models, real concurrency, fine-tuning on the same hardware, and (for firms) audit logs, multi-user auth, monthly compliance evidence, redundant power and a signed BAA. Solo builders get the big-model capacity for less than a cloud GPU rental; firms get the espresso-machine-to-commercial-kitchen jump. See the solo/developer plan →
Just building? Skip the call and start free → This form is for regulated firms that need a signed BAA/GLBA, dedicated GPUs, or on-prem — we reply within one business day.