Serve as the definitive three-way decision worksheet.
You will make a structured local, cloud, or hybrid decision and document privacy, cost, failure, and switching assumptions.
Before you begin
Describe one task, classify its data, and note hardware, connectivity, volume, budget, and maintenance ability.
Score the constraints
For the task, rate data sensitivity, required model capability, latency, offline need, volume, local hardware, maintenance capacity, and tolerance for provider change. Eliminate options that violate policy before comparing convenience.
Select the simplest fit
Choose local when a supported model meets the quality target and device control matters. Choose cloud when permitted data needs capability or scale unavailable locally. Choose hybrid only when splitting the workflow creates a clear benefit; otherwise it adds operational complexity.
Test total cost and reliability
Local costs include hardware, power, storage, and maintenance. Cloud services may charge by input, output, storage, tools, or provisioned capacity. Pricing and limits change, so link decisions to dated provider documentation rather than hardcoding assumptions. Test behavior during network or service failure.
Document the choice
Record which data crosses each boundary, who approves cloud transfer, how credentials are stored, what happens when a component fails, and how to switch providers. Revisit the decision when models, policies, or workloads change.
Compare the complete systems
Local can offer offline use and device control but needs compatible hardware, storage, updates, and troubleshooting. Cloud provides managed compute and broad model access but needs connectivity, account controls, and acceptance of current data and price terms. Hybrid can keep sensitive stages local while using remote capability selectively, but adds routing and consent complexity. None is automatically safest, cheapest, or fastest.
Three realistic examples
Private notes may favor a local model if it meets quality needs and the device is secured. Occasional analysis of public material may favor cloud access because no hardware purchase is required. A team might extract identifiers locally and send an approved redacted excerpt to the cloud with visible consent. These are examples, not rules; legal, policy, provider, and technical requirements can change the result.
Run a bounded comparison
Use five permitted examples and define success first. For local, record model, quantization, context, memory, responsiveness, and quality. For cloud, record model, settings, latency, quality, and estimated usage from current pricing. For hybrid, inspect the payload crossing the network and test connection loss. Date every estimate rather than treating hardware costs or provider prices as permanent facts.
Document and revisit
Record the selected architecture, rejected options, data boundaries, credential and update owners, spending controls, failure behavior, and test evidence. Keep prompts and sources portable. Define review triggers such as a new data category, model retirement, higher volume, policy change, or declining quality. A reversible decision reduces lock-in and turns future changes into deliberate work rather than emergencies.
Use gateways only when they clarify architecture
Decision: Prefer direct local or provider access for simple work; add routing only for a defined policy, portability, or observability benefit.
Relevant options: LocalAI and vLLM are self-operated serving approaches; LiteLLM and OpenRouter are different routing layers. Self-operation requires security and model-license review.
Worked example: Score one document task across local, cloud, and redacted hybrid paths for data, quality, latency, failure, maintenance, and dated cost.
Failure mode: Routing can obscure the provider, duplicate logs, add fallback transfers, or confuse cost ownership.
Validation: Trace endpoint and payload, disable unapproved fallback, simulate outage, and reconcile usage.
Useful next step: Choose the simplest compliant option and record when to revisit it.
Official sources checked 2026-10-24: LocalAI · vLLM · LiteLLM · OpenRouter. Product availability, support, terms, pricing, and licenses can change. These are vendor or project sources; no independent benchmark is claimed.
Original deployment decision worksheet
Download the worksheet. It covers task, data, capability, offline need, hardware, network, cost, owner, failure, export, and approval.
The synthetic example chooses local for a public, small classification task with offline need and compatible hardware. It is not a recommendation for private or regulated data.
Authoritative references
Reference scope checked 2026-10-27. These sources support the concepts and risk controls; product-specific behavior still requires the nearby vendor documentation and a dated test.
Your next step
Complete every field and choose the simplest eligible option; use hybrid only for a measured benefit.
Editorial provenance
Nyfir Studios is the publisher and editorially responsible organization. AI tools assisted with drafting and structured review for this candidate; factual claims, test labels, links, and retained evidence require human review before deployment. No named employee, credential, or unrecorded hands-on result is implied.