Choosing Hardware for DPF

Mac mini, Mac Studio, Framework Desktop, and DGX Spark configurations reviewed October 12, 2026; other machine examples last reviewed August 24, 2026. Product configurations, prices, and regional stock change frequently; follow the manufacturer links before ordering.

DPF has two customer deployment models:

1. Provider-assisted — the DPF application and business data run on the organization’s machine, while approved AI requests use an external model provider.
2. Local-first — the DPF application, business data, and primary language model run on the organization’s machine. A provider can still be configured as an explicit fallback for work the local model is not qualified to do.

Platform development is a separate workload. A customer should not have to buy a coding workstation to run day-to-day operations.

After choosing the host, continue with the Windows, macOS, Linux early-access, or single-VM cloud installation guide.

Linux pilot target

For the first Linux deployment, choose Ubuntu Server 24.04 LTS, x86-64, systemd and native Docker Engine with Compose v2, on a dedicated VM or physical server. Use the provider-assisted row below unless local inference is required. A GPU is optional for that profile. Docker Desktop is not needed.

This is a recommended pilot configuration, not a completed certification. Debian, Fedora, ARM64 and GPU configurations require their own acceptance results. The Linux guide tracks source fixes, release contents and the customer’s outstanding native checks separately. Avoid buying a GPU for this pilot before confirming the model, driver and container runtime together.

The short recommendation

The customer profiles size day-to-day operations. Embedded Build Studio can handle supported changes on a customizable install without separate Claude or Codex apps. It still needs a configured AI provider and development workspace; builds, tests, previews, source worktrees, and local coding models consume resources wherever they run. Size that additional workload separately. Removing external agent apps does not make a full development workload fit the operations profile.

A Mac mini for everyday business operations

A Mac mini is a suggested compact host for DPF as a business operating system: the application and business records run on the organization’s own machine, while people use the browser-based workspace to coordinate daily work.

• Mac mini with M6, 32 GB unified memory and a 1 TB SSD: matches DPF’s provider-assisted operations planning profile. Approved external AI providers handle model inference; those configured requests leave the host.
• Mac mini with M5 Pro, 64 GB unified memory and a 2 TB SSD: matches the practical Apple local-first profile. Qualify the chosen model, context, voice/document services, and simultaneous coworker work together.
• Mac Studio with M5 Max, 128 GB unified memory and a 2–4 TB SSD: consider this when larger local models, document-heavy workflows, or substantial development share the business host.

Apple currently lists a maximum of 32 GB on the M6 mini and 64 GB on the M5 Pro mini. The newer chip does not replace memory capacity, and a larger SSD does not replace RAM. See Apple’s Mac mini specifications.

These are workload-based suggestions, not an exact-model DPF certification or a promise that every business fits in 32 GB. The memory use of an actively developed 128 GB installation is not a measured minimum for ordinary operations. Before standardizing on 32 GB, verify a representative business workload, peak concurrency, memory pressure, restart recovery, and any enabled Build Studio work. Include a UPS and backups on separate storage with a tested restore path.

Local AI for professional firms

For a law firm that wants primary AI processing on-site, the M5 Max Mac Studio with 128 GB unified memory and a 2–4 TB SSD is a higher-capacity evaluation candidate. The memory leaves more room for the model, document processing, context, and the DPF services than either Mac mini configuration. Actual response times and simultaneous-user capacity need to be measured with the firm’s representative documents and chosen inference backend.

Apple lists 128 GB with the 18-core CPU / 40-core GPU M5 Max configuration. The M5 Ultra supports up to 512 GB unified memory; consider it only when measured model capacity or throughput justifies the larger machine. See Apple’s Mac Studio specifications.

Local hosting alone does not establish confidentiality or legal-work quality. Confirm which AI providers and integrations can receive documents, enforce matter access, test backup restoration, and qualify each intended AI workflow before using client material. An enabled external fallback can still send approved requests off-site.

Machines available now

These are representative, manufacturer-direct configurations available in the United States when this guide was reviewed. They are examples, not exclusive endorsements or DPF certifications. Match the specification rather than the logo, and include the warranty and on-site support appropriate for the organization.

Provider-assisted operations

There is little benefit in buying a large GPU for this deployment unless the organization expects to move primary inference on-site later.

Local-first Windows with NVIDIA

NVIDIA specifies 32 GB GDDR7, 575 W total graphics power, and a 1,000 W reference system-power requirement for the RTX 5090. Check the actual system’s power supply, cooling, clearances, warranty, and electrical circuit rather than assuming every 5090 tower is equivalent. See NVIDIA’s RTX 5090 specifications.

An existing RTX 4090 with 24 GB remains a capable DPF machine. It is not a reason to upgrade by itself. The 32 GB 5090 recommendation is for a new purchase because the extra 8 GB provides useful context and service headroom; it does not make the same model more knowledgeable.

Local-first Apple unified memory

Apple Silicon is a generally available DPF install surface. See the macOS installation guide. Memory cannot be upgraded after purchase, so buy for the expected model and useful life rather than today’s minimum.

Why unified memory changes the calculation

A conventional NVIDIA workstation has two memory pools:

• System RAM holds Windows, Docker, the DPF services, model mappings, browser work, and development tools.
• VRAM holds the actively offloaded model, context cache, and GPU compute buffers.

Putting a model in VRAM therefore does not eliminate its system-RAM footprint. It also means a 64 GB Windows workstation with a 32 GB GPU cannot present all 96 GB as one model-memory pool.

Apple Silicon lets the CPU and GPU share one high-bandwidth pool. DPF reserves 25% of unified memory for macOS and the application stack, leaving a planning budget of approximately 48 GB on a 64 GB Mac and 96 GB on a 128 GB Mac. That is why a 128 GB Mac can load models that do not fit a 32 GB graphics card.

Unified capacity is not the same as speed. The RTX 5090 has much higher graphics memory bandwidth and usually generates tokens faster when the entire model fits in 32 GB. Apple unified memory trades some peak inference speed for a much larger flexible memory pool, quiet operation, and excellent performance per watt. The right choice is therefore:

• NVIDIA 32 GB when low latency and CUDA compatibility matter most.
• Apple 128 GB when larger models, long context, portability, or a combined operations-and-development machine matter most.

Model and context sizing

Customer hardware should be sized for a tool-using operational model, not a coding benchmark. DPF model qualification prioritizes correct tool selection, valid arguments, structured output, instruction following, authority boundaries, and recovery after a rejected call.

As a useful current planning envelope, OpenAI’s gpt-oss-20b is a 21-billion-parameter, 3.6-billion-active-parameter open model with function calling and structured outputs. The llama.cpp project’s deployment measurements put it at approximately 14.9 GB for 8K context, 15.5 GB for 32K, and 17.9 GB for 131K. A 16 GB GPU can technically approach this class of model, but leaves too little operating margin to be the standard customer recommendation.

DPF uses 32K as the practical buying baseline and retrieval tools to bring the relevant records into the prompt. Larger windows are useful for specific workflows, but they increase memory use and latency and do not automatically improve tool judgment.

For several simultaneous AI conversations, use an approved provider or a throughput-oriented inference service. Docker recommends its llama.cpp engine for memory-efficient local and single-user work, and vLLM on supported NVIDIA hosts for higher throughput. See Docker Model Runner inference engines.

Compact Linux candidates — qualification required

Linux runs DPF containers through native Docker Engine without Docker Desktop’s additional VM. That can reduce hosting overhead; DPF has not measured a general efficiency advantage across these machines. The Linux install guide remains early access, with Ubuntu 24.04 x86-64 as the recorded first pilot target.

Framework Desktop is the PC alternative to evaluate when a shared memory pool and a choice of Linux or Windows matter. DGX Spark is the smaller NVIDIA local-AI candidate. Neither is a qualified DPF replacement for the generally available Apple Silicon path yet. For provider-assisted operations, a conventional compact x86 business PC with 32 GB RAM and a 1 TB SSD can instead be sized against the operations profile, with Linux pilot acceptance still required.

Docker Model Runner lists CPU, NVIDIA CUDA, AMD ROCm, and Vulkan backends for Docker Engine on Linux. Backend support does not establish support for every GPU or model. Do not carry a Linux AMD recommendation over to DPF’s Windows local-inference path without separate qualification.

Before recommending either candidate for a business, verify the exact DPF installer, GPU backend, model, tool-call parser, peak concurrent workload, reboot, upgrade, backup, and restore path. The same validation applies to a law firm’s local document workflows; hardware capacity alone does not establish confidentiality or legal-work quality.

Before ordering

• Prefer Windows 11 Pro or a currently supported macOS release.
• Buy at least 1 TB storage; choose 2 TB for local models and 4 TB for contributor work. Keep backups on separate storage.
• Include a UPS for a machine that hosts the organization’s operational data.
• Check noise, thermals, electrical service, warranty, and replacement time.
• Treat GPU memory, system memory, model format, context length, and concurrent services as one capacity plan.
• Recheck this dated guide and the live manufacturer configuration before purchase.