PRIVATE INFERENCE / HARDWARE + RUNTIME
On-prem Mac Studios and CueCloud’s custom inference stack, configured for your models and team. Aether-1, our proposed Memory Processing Unit, is the next hardware generation in development.
Plan your Vault deployment ↗Consumer-grade hardware, sized for your models and installed by our team.
System concept / configuration varies by deployment
VAULT / THE HARDWARE EVOLUTION
Gen 1 uses Mac Studio with our inference software. For Gen 2, we’re developing Aether-1, our own Memory Processing Unit.
The same memory serves CPU and GPU. Serving capacity depends on the model, context and concurrent workload.
CURRENT DEPLOYMENT PATH
Apple Silicon Mac Studios, installed on-site with CueCloud’s custom inference stack. We configure the hardware, models and local serving around your team’s workloads.
The CPU and GPU access a common memory pool. Our inference software manages model execution, memory and scheduling within the hardware’s capacity and bandwidth.
Gen 1 runs on Apple Silicon. For Gen 2, we’re developing both the processor and the software that runs on it.
GEN 2 / WHY WE’RE BUILDING AETHER
Running AI on your premises means working with the power, cooling and space you have. We’re building Aether to run larger models within those limits.
When a model exceeds an accelerator’s memory, adding devices can become necessary just to hold it. You pay for the extra compute, then supply the electricity and cooling it needs. We think inference hardware should let you add model capacity without tying every increase to more compute.
More room for models should not require the same increase in compute hardware.
01 / CAPACITY
Model weights must be available when a request needs them. Keeping more weights in memory does not always require proportionally more computation. We want to use that distinction to support larger models and keep several models available on-site.
02 / BANDWIDTH
The processor needs a steady supply of model weights and request data. If memory cannot supply them fast enough, answers slow down. Capacity and bandwidth have to grow together to make larger models practical to use.
03 / POWER
An office or warehouse has a power and cooling budget. That budget must cover the computers, memory and networking needed to serve your team. We’re designing Aether around those costs and will measure them across the whole system.
Mac Studio gives our inference software a memory pool shared by the CPU and GPU. That is the basis of Gen 1. Software can improve how we use the machine, but it cannot change how much memory it has or how quickly the hardware can read it.
With Aether-1, we can choose the balance of memory capacity, bandwidth and computation ourselves. We’re developing the processor and its software together for local inference, with larger models and lower deployment power as our goals. The implementation is proprietary.
Aether-1 is in development. We still need to test capacity, response speed and power together under realistic workloads. Results will depend on the model, context length and number of simultaneous requests. We have not announced product specifications or a release date.
Vault / INCLUDED IN THE DEPLOYMENT
Hardware, software, installation, and handoff.
Compact inference hardware, sized for your models and team. We plan the power, networking, and installation with you.
Compute and memory matched to the models you want to run, the context they need, and the number of people using them.
We work through available power, cooling, physical placement, and local networking before installation.
The hardware arrives as part of a configured inference deployment, with the supported configuration agreed up front.
A configured local inference system sized for your team and ready for installation.
MODEL SERVING / BUILT INTO VAULT
Vault is built for inference: running trained models to answer requests from your applications. We supply the hardware and our complete inference stack, licensed to your organization and configured as one system.
YOUR APPLICATIONS / BLITZ / INTERNAL TOOLS
Model selection · Installation · Validation
Model execution · Memory management · Scheduling · Local serving
Our runtime is part of your Vault deployment. You receive a license to use it under your agreement; model weights retain their own license terms.Compact footprint → More model capacity → More concurrent work
START WITH WHAT YOU WANT TO RUN
Code navigation, proposed changes, tool calls, and test feedback put different demands on a model. We help select and configure models for the way your engineers and agents actually work.
We install the agreed model artifacts, configure supported formats and quantization, set context and concurrency limits, and connect your applications to the local endpoint. Then we validate the configuration on representative tasks with your team.
Choose hardware around a focused workload and the models that fit its memory budget. We tune the serving configuration for your team’s expected use.
Model weights, working context, and runtime overhead all need memory. We size them together and assess supported hardware arrangements before committing to a configuration.
More simultaneous requests can call for a different compute layout or separate serving capacity. We customize the deployment around demand, latency goals, power, and placement.
Bring us the models you want to run, or the tasks you need them to do. We help choose the configuration and set it up. Training and fine-tuning are separate requirements to scope.
01 / THE PHYSICAL FOOTPRINT
We check power, airflow, placement and service access before installation.
Illustrative enclosure and airflow. Final equipment layout and clearances follow the selected hardware.
We review the selected hardware, available power, operating schedule, and any UPS requirements with your facilities team. We account for both average consumption and peak demand.
EXPLORE THE FOOTPRINT
Adjust an example load. These are planning inputs, not Vault specifications or measured performance.
Energy = units × average watts × operating hours × 30 ÷ 1,000. Electrical input becomes approximately the same heat load in the room. Excludes idle time outside the selected hours, networking, UPS losses, and room cooling. Circuit and cooling sizing also require peak-load and site information.
02 / YOUR OPERATING BOUNDARY
Choose the deployment pattern closest to your environment. We work with your IT and security owners to define the actual connections, dependencies, and access paths.
INTERNAL SERVICE
Your engineering tools and internal applications reach Vault through the network paths your team approves. We scope endpoint access, identity integration, and operational dependencies with your IT owners.
We document the approved paths and validate them with your system owners during setup.
03 / SETUP IS PART OF THE SYSTEM
You should not have to assemble a model-serving project from parts. Our team handles the agreed setup, with your facilities, IT, and workload owners involved where their environment requires it.
You bring representative tasks, preferred models, expected users, and the zone where the system will run. We turn those into a hardware and deployment plan.
We assemble the selected configuration, install the inference stack, and configure the best-fit open-weight models for its compute and memory budget.
We coordinate with your IT and facilities owners, connect the system to the agreed power and network, and configure the endpoint and application connections.
We exercise your representative workload, check operating behavior, and walk your team through everyday use, restart procedures, and the agreed support process.
AFTER INSTALLATION
We agree how model updates, maintenance, incident triage, and support access will work. In a restricted or offline zone, those procedures are part of the deployment design.
THE COMPLETE SYSTEM
Vault runs the models. Blitz puts them to work.
Your organization controls the environment.
Your engineers work in Blitz. CueCode IDE is included.
Run the models behind Blitz on hardware inside your environment.
For an air-gapped installation, model access, identity, updates, and support are configured and validated within the agreed boundary.
Tell us about your team and workload. We’ll define the hardware, software, and support scope with you.
Start a conversation ↗