On-Premise & Offline AI Deployment Guide Three degrees of offline, twelve design tests, and four patterns. A design method, not a model bake-off, not a hardware calculator, and not a certification. Edition: November 2026. 21 pages. Written for: Security, infrastructure, operations and technology leaders who have to decide whether a workload can leave, and the sponsors who have to fund what follows. Canonical URL: https://asymilinkmeta.com/resources/offline-ai-deployment-guide Written for the search query: on-premise offline AI deployment guide Key takeaways - “Offline” is always used with a degree — D1, D2 or D3 — or it is not yet a design. A private endpoint is not automatically any of them. - Twelve design tests, including a map of what would leave, and hosting decided before the case is funded. - Four patterns: inference only, retrieval fully local, local tools, and air-gapped operations. - Sizing is a design, with questions and illustrative classes. This guide will not publish a SKU, a memory figure, or a tokens-per-second claim. What's in this guide This 21-page guide is a technical design method for deciding whether a workload can leave, and for designing the work that follows. The library has sometimes described it as a model-selection guide, a quantization bake-off, and a hardware calculator. It does not contain those things. We will not rank foundation models by name, we will not declare a quantization method “best,” and we will not publish a SKU, a memory figure, or a tokens-per-second claim — invented or otherwise. What it is, instead, is three degrees of “offline,” twelve tests to apply before a case is funded, a way to size a workload with questions and labeled illustrative classes, and four patterns for inference, local retrieval, local tools, and air-gapped operations. AsymiLink AI — our in-house model and inference platform — runs 70B+ and 400B+ parameter models on the customer’s hardware, with an optional air gap. Those model sizes are a product capability, not a default of every engagement. We are vendor-neutral: a recommendation can be AsymiLink AI, a stack you already run, or no build at all. It is written for security, infrastructure, operations and technology leaders, and for the sponsors who have to fund what follows. This guide is not legal, security or regulatory advice, and it is not a certification that a system is offline, air-gapped, or authorized for any regime. A companion workbook holds a design-test register, a degree-of-offline sheet, a sizing questionnaire, and a pattern register. A completed workbook is not a certification. Name the degree, then apply the tests Teams use one word for three different designs. A private endpoint, a contractual promise not to train, and a rack with no path out are not the same thing. D1 is no training telemetry. D2 is no inference egress. D3 is a full air gap. Each step is a tighter claim. Moving right is a design decision, not a marketing upgrade. The continuum is not a certification scale and not a ranking of products. A team that says “offline” and means a private endpoint has not finished T1. Hosting is decided before the case is funded. A pilot that has already sent data to a hosted tool “to see if it works” has already taken the decision. What would leave is mapped: prompts, completions, embeddings, logs, traces, evaluation sets, crash reports, license checks and training telemetry. The model and the corpus stay in the approved environment; a “temporary” copy on a laptop, or a vendor evaluation, is still a copy. Updates do not become a standing path out. Monitoring does not phone home, or that egress is named, limited and approved. Sizing is a design, not a shopping list — questions and illustrative classes, no SKUs and no invented throughput figures. Quality is measured against a frozen set, including after a quantization or a model swap. Downstream tools are in view: any tool that becomes a path out is a T1 and T2 question again. Training and reuse are written down. Companion obligations are asked, not assumed away: an air gap is not coverage of a design that was never finished, and it is not a certification. Three degrees of offline “Offline” is always used with a degree, or it is not yet a design. A private endpoint is often none of these. D1. No training telemetry. The inference path may still leave. A contractual promise not to train is a D1 claim only if what would leave is mapped and the promise is the contract. D2. No inference egress. Prompts, completions and the working path stay in the approved environment. This is tighter than D1, and it is still not a full air gap. D3. A full air gap. No path out, including updates, license checks and monitoring, unless a named, limited exception is approved. This is not a certification. The author’s design continuum. It is not a certification, and it is not a ranking of products. T1 is the test that names the degree — or writes a reason that a hosted path is acceptable. Four patterns (excerpt) An illustrative sequence. It is not a required path and not a delivery timetable. The honest first case is often P1 or a read-only P2. P1. Inference only. A local model answers a prompt. There is no retrieval corpus beyond what the person put in, and no tools that call other systems. A drafting assistant, with a person who can stand behind what leaves the room, is often an honest first case. P2. Retrieval fully local. A question is answered from a corpus the organization already holds. Embeddings, retrieval, the corpus and generation all sit in the approved environment. The RAG architecture guide is the companion for citations, permissions, chunking and evaluation. P3. Local tools. Tools the agent can call stay in the approved environment. Any tool that becomes a path out is a T1 and T2 question again. The agentic-workflows paper is the companion for the gate. P4. Air-gapped operations. The workload, the updates, the monitoring and the trail have no standing path out, or a named, limited exception is approved. Isolation does not make an operator or a switch-off optional. Twelve design tests sit under every pattern, including T8 (sizing is a design, not a shopping list) and T9 (quality is re-measured after a change). The complete tests are in the PDF and the workbook. Frequently asked questions Q: Is a private endpoint “offline”? A: Not automatically. T1 asks you to name the degree — D1, D2 or D3 — or to write a reason that a hosted path is acceptable. A team that says “offline” and means a private endpoint has not finished this test. Q: Does this guide recommend a model, a SKU or a quantization method? A: No. It will not rank foundation models by name, declare a quantization method “best,” or publish a SKU, a memory figure, or a tokens-per-second claim. Sizing is a design, with questions and illustrative classes. Q: Is this a certification that a system is offline or authorized? A: No. This guide is not legal, security or regulatory advice, and it is not a certification that a system is offline, air-gapped, or authorized for any regime. An air gap is not coverage of a design that was never finished. Q: What is in the companion workbook? A: A design-test register, a degree-of-offline sheet, a sizing questionnaire, and a pattern register. A completed workbook is not a certification. This page is the public summary. The full PDF and companion workbook are available through the request form on the page. There is no public file URL that bypasses that form.