New high-performance hardware>$ 100KAdditional cost
Scale your local AI
beyond one device.
Connect consumer hardware you have already paid for into one distributed inference runtime.
Zero or lowest new hardware cost
Enhance local AI
at lowest cost.
No need to buy a higher-performance workstation. Pool the heterogeneous devices you already own—laptops, desktops, tablets, Mac Minis, and devices of different operating systems.
With PRIMA: collect/add more devices$ 0–10KLowest cost · useful is enough.
Scale what matters
Scalability,
not just a speed miracle.
Bring more of your existing hardware together to expand what local AI can do—without pretending one device becomes something it is not.
Application cases
Bring every device into your local AI.
Your devices. Your private AI.
Pool the hardware around your home into one private inference service.
Shared capacity for the team.
Turn existing office devices into a useful local AI pool for everyday work.
Different LANs. One AI pool.
Pool devices across remote LANs and strengthen the local AI you use here.
Flexible by design
Easy to deploy. Free to adapt.
PRIMA manages devices as they join and leave, without imposing a specific device or network setup.
Devices join automatically.
When a device comes online, PRIMA discovers it and adds it to the inference pool—no complex configuration required.
Devices leave. PRIMA adapts.
When a device goes offline, PRIMA detects its departure and heals the inference pool automatically.
Use the devices and network you have.
PRIMA has no specific hardware or network requirement. Mix heterogeneous devices across wired LAN, Wi-Fi, or secure cross-LAN connections—including different device types, operating systems, chips, memory, GPUs, and storage.
Prima Lab · Blog
Notes on practical local AI.
The workstation you need may already be in the room.
Local AI does not always need another monolithic machine. Sometimes the more practical path is to treat the hardware around you as one adaptable pool.
Local AI often begins with a shopping list: more VRAM, a larger power supply, and a workstation built around the biggest model you hope to run. That approach can be fast, but it also leaves capable laptops, desktops, mini PCs, and tablets sitting outside the system.
The single-box assumption
A single workstation is easy to reason about because every resource sits behind one operating system. Its limits are equally clear. Once the model, context, or precision outgrows that box, the usual answer is to replace it with a more expensive one.
A useful local AI system does not have to be the fastest possible system.
Treat hardware as a pool
Distributed inference changes the unit of planning from one machine to a collection of machines. A laptop can contribute a GPU, a desktop can contribute memory and compute, and devices on another LAN can remain part of the same testbed.
The important question becomes less binary: not “Can this device run the model?” but “What useful capacity can this device add to the pool?”
Useful is a product decision
A pooled system may not beat a purpose-built server on every latency metric. It can still unlock a larger model, longer context, better precision, or enough throughput for a private workflow—without discarding hardware that has already been purchased.
That tradeoff is the point: local AI should be able to grow incrementally, across the devices and networks people actually have.
A shared runtime
for the lab.
Turn mixed CPU and CUDA machines into a single inference service.
Resource-aware
Edge intelligence,
kept close.
Serve apps and agents from infrastructure you control, near the work and the data.
Self-hosted endpoint
One runtime.
Many devices.
Continuous inference.
Tokens move across assigned CPU and GPU resources. Membership changes are applied between requests.
- Active token flow
- Standby path
- Node joining
- Node leaving
Prompt input
Explain how transformers work.
LaptopCPU
WorkstationCPU + GPU
Mini PCCPU
GPU nodeCUDA
Token output
Example token output: Transformers use attention to model.
Old laptopLeaving after request
New mini PCJoining next request
OpenAI-compatible endpoint
Collaborate across
locations. On hardware
you control.
Coordinate inference over operator-provided reachable networks while execution stays on operator-controlled devices.
Lab LAN
Operator-controlled devices
Reachable path
Office LAN
Operator-controlled devices
Playground
Configuration-specific measured-rate playback.
Useful AI capacity
is already around you.
Build from compatible hardware that has already been purchased.
01
Use what you have.
Laptops, desktops, mini PCs and consumer GPUs become useful capacity.
02
Keep data local.
Prompts and model execution stay on infrastructure you control.
03
Resilient collaboration.
Eligible devices can join, leave and recover at controlled boundaries.
Evidence you can inspect
Built in the open.
Measured on real hardware.
- ICLR
2026 - Published evaluation
- 80+
- GitHub stars
- MIT
- Licensed
- 674ms / token
- 70B TPOT · four-device home testbed
- <6%
- 70B · per-device memory overhead
Active development
Additional large-model and MoE evaluations are underway. Results will be published with reproducible artifacts when they are ready.
Published results depend on the cited model, quantization, hardware, networking and workload; they are not universal performance claims.
Start with Prima.cpp
Build private intelligence
where your data lives.
Open source. Self-hosted. Built for heterogeneous hardware.