Fit, not just VRAM
Weights + quant overhead + architecture-aware KV cache + workspace + safety margin.
Open hardware decision system · US market
Build the right local AI machine.
Choose your workload, budget, context, and number of agents. ModelRig tells you what to buy, what it will actually run, and how to use it remotely.
Decision trace / catalog 2026.08.18-1
Data as of 2026-08-18
Data as of 2026-08-18
01 · workloadagents × context
02 · memoryweights + KV + margin
03 · evidencerange + confidence
04 · outputbuy + deploy
Inverse profiling
Start with the work. End with a bill of materials.
A 32 GB card is not automatically a 32 GB solution. ModelRig tests the complete working set, runtime, circuit, form factor, and multi-agent load before scoring a system.
Ranges, not theater
Measured when exact. Interpolated when comparable. Wide spec-derived bands otherwise. Unknown stays unknown.
Operations included
Power, acoustics, topology, reliability, remote access, and ownership cost affect the ranking.
One shared engine
The web app and installable CLI execute the same deterministic TypeScript package.
Buying guide
Five materially different tiers.
Used-value, efficient new, high-bandwidth, multi-GPU, and professional-memory systems. No cosmetic variant padding.
Entry
Compact used RTX A2000 workstation
A quiet, low-power entry point for 7B coding models and remote always-on use in very limited space.
ValueEfficient RTX 5060 Ti 16 GB tower
Low operating cost and mature CUDA support for responsive small models, with 16 GB allowing constrained 24–30B Q4 use.
PerformanceUsed RTX 3090 24 GB tower
High bandwidth and 24 GB VRAM make a used 3090 a strong price/performance step for 24–32B Q4 models at moderate context.
Multi-agentDual used RTX 3090 agent workstation
Two independent 24 GB GPUs are useful for concurrent agents or supported tensor parallelism, with enough PCIe lanes and spacing to avoid a consumer-board trap.
No-compromiseRTX PRO 6000 96 GB workstation
A single 96 GB ECC GPU avoids consumer multi-GPU pooling pitfalls and supports 70B-class models with substantial long-context and concurrency headroom.
Remote by design
Your machine, reachable from anywhere—not open to everyone.
Generated bundles use a private Tailscale overlay, authenticated SSH tunneling, loopback-bound APIs, health checks, and automatic restart. No router port forwarding.
Private network
Restrictive device ACLs establish identity before the host is reachable.
SSH tunnel
The inference port remains on loopback; clients forward it through authenticated SSH.
Inspectable service
Compose or LaunchAgent files are generated as text. Privileged package changes are never hidden.