For companies that cannot send documents, code or customer data to a public AI service. We size the hardware, install and tune the models, put the agent software on every seat, and stay for support.
vLLM or Ollama on an RTX 4090, RTX 5090 or DGX Spark, or on GPUs you already own. Tuned for your team size and benchmarked with your own prompts before go-live.
The desktop agent for business users, pointed at the office model server. Data stays in one local database file per machine; no accounts to manage.
A terminal coding agent for every engineer, with the four-agent team mode, skills and MCP, using the same model server.
Onboarding sessions, a shared MCP and skills setup, model updates, software upgrades and a support line.
Memory decides which model fits; concurrency decides how many people it serves well. The table is our starting point, and the sizing step replaces it with your numbers.
| GPU | Memory | What fits | Team | Basis |
|---|---|---|---|---|
| NVIDIA RTX 4090 | 24 GB | Models up to about 14B parameters at full speed; larger ones quantized | A small team doing documents, email and chat | Guidance |
| NVIDIA RTX 5090 | 32 GB | Qwen 3.8 27B in NVFP4 on vLLM, long sessions with a CPU cache tier | 8 developers per card for interactive coding agents; 16 workable | Measured |
| NVIDIA DGX Spark | 128 GB unified | 70B-class models and long context on one desk-side box | Larger teams, or bigger models for fewer people | Guidance |
Measured with Nucleus agents on one RTX 5090: 8 developers at about 57 tokens per second each with all sessions resident; 16 developers all finish their tasks at roughly 1.5 times the time.
Nucleus benchmark, 2026-09-01/02, Qwen 3.8 27B NVFP4 on vLLM 0.28
We size per team from your prompts, context lengths and headcount. Existing GPUs and multi-GPU servers are welcome.
For regulated or confidential work. Public egress can be blocked at the firewall.
Stays inside
Leaves the network
Local by default, with hosted models on Nucleus AI Cloud for approved people and workspaces, through the Nucleus gateway.
Stays inside
Leaves the network
These are the properties of the shipped software. Your security team can verify each one on a test machine before rollout.
Nucleus Work stores users, sessions and messages in a single SQLite file on the machine. No Postgres, Redis or cloud database.
The desktop app starts straight into the workspace. Cloud linking exists but stays off unless a user links an account.
Every agent module runs as a WebAssembly plugin inside the host, talking through a narrow, JSON-only interface.
Model keys are written with private file permissions and referenced by path. They are never placed in a config file.
Apps that Nucleus Work builds preview on 127.0.0.1 of the same machine. Nothing is exposed to the LAN or the internet.
Most deployments take weeks, not quarters. You keep the hardware, the models and the data; we do the sizing, the installation and the tuning.
A short workshop on the workflows you want automated, the data involved, your network policy and how many people will use it.
We propose the GPU, the model and the deployment shape for your headcount. Hardware you already own is fine.
We set up the inference server, load the models, benchmark with your prompts, and put Nucleus Work and the CLI on every seat.
Team onboarding, model updates, software upgrades and a support line for as long as you want one.
Not for day-to-day use. Fully offline installs run with public egress blocked. Internet is only needed if you choose the hybrid option, and then only for the workspaces that use a gateway model.
Yes, in hybrid mode. An admin connects the subscription your company owns to Nucleus AI Cloud, sets per-member quotas, and the hosted models appear in the picker in Nucleus Work and the CLI.
Open-weight models served with vLLM or Ollama. Our current reference is Qwen 3.8 27B on an RTX 5090. We choose per team from your prompts, languages and context length.
Yes. We assess it in the sizing step. Anything with an NVIDIA GPU and enough memory for the chosen model is a candidate, and multi-GPU servers work with vLLM.
That is a scoped project rather than an off-the-shelf install. If you need the hosted platform experience inside your network, we discuss it at the assessment and quote it separately.
Singapore. We work on site in Singapore and remotely across the region.
A few lines about what you want automated, roughly how many people would use it, and whether you already have GPUs. We reply with a proposed shape and next steps.