Private AI for your code, documents, and daily work—running on hardware you own.
Ownership / 01
We select and install leading open-weight models for the hardware you buy. You choose what runs. There is no per-token charge for the models on your box.
Alfred Remembers keeps conversations, documents, recordings, photos, and notes on the workstation. Local inference and document search can run without sending prompts or files to an external model provider.
The computer is yours, set up for the way you work. Offline configurations are available. Network access is only for the connections you enable, such as web research, email, or software updates.
Alfred Remembers / 02
Alfred Remembers connects the conversations, documents, recordings, photos, and notes you add into a lasting memory stored on your Intelligence Box.
Ask “What did we decide about that contractor?” or “What was the restaurant Sarah recommended?” Alfred helps find the relevant history, so you can pick up where you left off.
For families, it brings together household documents, purchases, warranties, preferences, and recommendations. For businesses, it connects meetings, projects, customers, procedures, and decisions.
Your memory lives on hardware you control and can be used by supported local or private AI models. It is included with every box.
The Boxes / 03
Each one is hardware you own, models selected for that hardware, and our software—installed and ready for a kind of work.
Solo
A compact Mac mini with 64 GB of unified memory, configured for one person using a coding agent, private chat, and document search. A practical starting point for local AI, with compact models selected for useful quality and responsiveness.
Apple M5 Pro · 18-core CPU / 20-core GPU · 64 GB unified memory · 1 TB SSD
Pro
A Mac Studio with 128 GB of unified memory for larger coding models, longer working context, and heavier document tasks. Built for a developer or power user who wants more model choice and room to grow.
Apple M5 Max · 18-core CPU / 40-core GPU · 128 GB unified memory · 1 TB SSD
Office
Two Mac Studios in our custom enclosure, connected over Thunderbolt 5 and your local network. Separate model services keep coding work and everyday business requests from competing for the same machine. Designed for a mixed team of up to eight people, subject to workload validation.
2 × M5 Max Mac Studio · 128 GB unified memory and 1 TB SSD per node · 4 TB external document storage · 10 GbE
Team Pro
A Linux workstation with two 96 GB NVIDIA professional GPUs for demanding coding agents, shared inference, and CUDA workflows. Designed for an eight-person team with one or two active coding agents alongside business users, with workloads and context limits agreed during setup.
2 × RTX PRO 6000 Blackwell Max-Q · 96 GB GPU memory each · Threadripper 9960X · 256 GB ECC · 8 TB NVMe
| Solo | Pro | Office | Team Pro | |
|---|---|---|---|---|
| Best for | One person | A developer or power user | A small mixed team | Sustained team AI work |
| System | Mac mini | Mac Studio | 2 × Mac Studio in a custom enclosure | Linux workstation |
| Processor | M5 Pro, 18-core CPU | M5 Max, 18-core CPU | 2 × M5 Max, 18-core CPU | Threadripper 9960X |
| Graphics | 20-core GPU | 40-core GPU | 40-core GPU per node | 2 × RTX PRO 6000 Blackwell Max-Q |
| Memory | 64 GB unified | 128 GB unified | 128 GB unified per node | 192 GB GPU, plus 256 GB ECC |
| Storage | 1 TB SSD | 1 TB SSD | 1 TB SSD per node, 4 TB external | 8 TB NVMe |
| Network | 2.5 GbE | Built-in Ethernet | 10 GbE and Thunderbolt 5 | Confirmed with the build |
| Software | MLX stack | MLX stack | MLX, separate model services | CUDA stack |
| From | $4,495 | $7,495 | $15,995 | $59,995 |
Team size describes the intended mix of users, not an unlimited number of simultaneous generations. Supported models, context lengths, and concurrent workloads are confirmed for each configuration. Apple unified memory and NVIDIA dedicated GPU memory have different capacity and performance characteristics. Multi-node memory totals are distributed across machines. Prices are U.S. turnkey prices before tax and shipping.
Workflows / 04
Explore a repository, make changes, generate tests, and run approved tools using a locally hosted coding model. Your source code can stay on your network.
Search selected files and receive answers linked to source passages. We connect your documents through retrieval, so you can update the knowledge base without retraining the language model.
Extract information, organize reports, draft explanations, and work with spreadsheet or calculation tools. Review important outputs and use the underlying tools to verify numbers.
Draft briefs, proposals, internal communications, and marketing content. Use local files privately, and enable external research or connected services when your workflow calls for them.
Software
We select and pre-install leading open-weight models, configure the tools for your workflows, and provide our own software, including Alfred Remembers. Team systems add user access, request routing, workload queues, and monitoring.
Apple systems use an Apple Silicon inference stack based on MLX and compatible local runtimes. NVIDIA systems use a CUDA inference stack selected and tested for their models.
Our custom Apple enclosures organize the computers, cabling, and airflow while keeping individual nodes accessible for service. Multi-Mac systems can run independent model services or supported distributed workloads over Thunderbolt 5.
Why local
Local inference and document retrieval can operate without sending prompts or files to an external model provider.
You choose the models and integrations. Connections that need the network are explicit during setup.
No per-token charges for the models running on your box. The hardware is a purchase, not a meter.
The system is configured for coding, documents, writing, and operations—not a generic cloud dashboard.
Once the selected models and software are installed, the box can work without a network.
You keep the hardware. Optional hosted models stay available for work that benefits from them.
Models
Our current evaluation set is below. We select the model and quantization for your hardware and work, and qualify new releases before adding them to the supported catalog.
Larger models and longer context windows need more memory and may reduce concurrency. We demonstrate the configured system on a representative workflow so you can judge its usefulness before choosing.
FAQ
Alfred Remembers connects the information you add into a persistent local memory that supported AI models can draw on. It helps recover the context behind a question across conversations, documents, and other saved material.
It provides a familiar chat experience using locally hosted models. Quality and features depend on the selected model, tools, and configuration. We demonstrate your intended workflows rather than promise identical results to every hosted assistant.
Yes, once the selected models and local software are installed. Workflows that depend on online information or cloud applications need network access.
Yes. Team configurations provide shared access and managed queues. We size the system around how many people generate responses at once, their context needs, and how heavily they use coding agents.
Supported distributed software can split a model across Macs. Each Mac still has its own memory, and distributed execution has communication overhead. For many office workloads, running separate model services offers a better experience.
It can replace the workloads that your local models and tools handle well. Savings depend on actual usage, hardware cost, electricity, and support. Optional hosted models can be configured for work that benefits from them.
Multi-Mac systems can expand with additional qualified nodes. GPU workstations have expansion paths determined by their motherboard, chassis, power, and cooling. We document the supported upgrades with your configuration.
Contact
Tell us who will use it, what they will do, and which data needs to stay local. We'll recommend a box and demonstrate a representative workflow.