--- title: "A LoRA Adapter Is Not the Whole Model: Compare the Delivered AI Artifact" description: "Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20." canonical: "https://www.leviathansystems.co/articles/lora-adapter-versus-merged-model-delivery" last-updated: "2026-10-20" --- # A LoRA Adapter Is Not the Whole Model: Compare the Delivered AI Artifact Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20. A small LoRA checkpoint can be a complete adapter while leaving the base model to be supplied separately. Choose between adapter-plus-base and a supported merged model by comparing the exact retained artifacts, dependencies and task-selection capabilities. Download size alone cannot show whether the delivery covers the task-specific model you intend to serve. ## Separate the base, the learned update and the inference delivery The original LoRA paper's section 4.1 describes a pretrained weight matrix W0 and an update represented as the product BA. W0 stays frozen during adaptation; the smaller matrices B and A contain trainable parameters. The adapted layer uses W0 plus BA, with the applicable scaling. The learned update changes how the base operates without replacing the need for its parameters. Keep three objects visible in a delivery discussion: the original base parameters, the learned update and the complete supported artifact package for inference. The third may retain the first two separately or incorporate the update into a merged model. This differs from reducing the bit width of stored weights, the decision covered by the related quantization guide. It also differs from retaining optimizer and progress state to continue training, covered by the checkpoint-versus-snapshot guide. Here the decision is how a task-specific adapted model is delivered for serving. References: [R1](https://arxiv.org/html/2106.09685v2) ## Count a smaller update without deleting the base Consider a deliberately simplified single layer with a 1024 by 1024 base matrix and an invented LoRA rank of 8. The base matrix contains 1,048,576 scalar entries. B has dimensions 1024 by 8, containing 8,192 entries. A has dimensions 8 by 1024, also containing 8,192 entries. Together, the two trainable matrices contain 16,384 entries. Dividing 1,048,576 by 16,384 gives a 64-to-1 entry-count ratio between this base matrix and its two update matrices. The product BA has the dimensions needed to update W0, but its factors can be retained without storing that full product separately. The base still exists. This original arithmetic omits scaling and fixed metadata from the count, and rank 8 is an illustration rather than a recommendation. It establishes no full-model compression, VRAM, training-time, quality or speed ratio. It does explain how an update can be much smaller than the matrix it modifies. References: [R1](https://arxiv.org/html/2106.09685v2) ## Read an adapter delivery as a dependency on an exact base Hugging Face's PEFT checkpoint guide distinguishes adapter weights from adapter configuration. Its adapter weight file omits base-model weights; the configuration describes the adapter and can identify the base model and revision. A checkpoint can therefore be complete for its adapter purpose while leaving the inference delivery incomplete. Ask the supplier to resolve the exact base revision, including when a configuration leaves the revision unspecified. A model-family name or a reference that can move does not give the buyer a fixed identity to retain. Matching tensor dimensions are insufficient evidence that a different base revision is suitable. In the simplified layer, substituting different values for W0 changes the sum even if BA still fits its shape. The supplier needs to declare the supported pairing; the buyer should not silently substitute a newer base. References: [H1](https://huggingface.co/docs/peft/developer_guides/checkpoint) ## Identify what the supported merged form retains A supported merged full-model delivery contains the base weights with the learned update applied. The separate base-weight download is no longer needed to complete that delivered weight set. Configuration, tokenizer and runtime requirements still belong in the inference package description. The PEFT checkpoint guide warns that merging is unavailable for some methods or settings, including certain quantization configurations. Its merged-and-unloaded form also loses PEFT adapter controls such as disabling an adapter or loading multiple adapters. PEFT's v0.21.0 LoRA reference distinguishes that standalone result from a merge that retains adapter machinery for later unmerging. Ask which form the supplier supports and delivers. The word merged alone does not define the controls you retain, guarantee a reversible or lossless conversion, or establish inference performance. References: [H1](https://huggingface.co/docs/peft/developer_guides/checkpoint), [H2](https://huggingface.co/docs/peft/v0.21.0/package_reference/lora) ## Compare one base with two adapters against two merged variants Consider a fictional buyer with two tasks: extracting specified fields and summarizing a document. The example stipulates one exact base, B1, and two distinct, compatible task adapters, A1 and A2. These labels name whole adapter artifacts, not the individual A matrix in the layer example. No task-quality result is assumed. An adapter delivery retains B1 once, together with A1 and A2 and their required configuration. For this comparison, that is one base and two update artifacts. If the supplier supplies only A1 and A2, the offer must identify who supplies and retains B1. A separate supported arrangement supplies M1, the complete merged variant for B1 with A1 applied, and M2, the complete merged variant for B1 with A2 applied. That is two complete merged variants. Each incorporates the base weights needed by its variant. These are artifact populations. They are not necessarily three versus two files, running model copies, GPUs or concurrent services. Packaging and the supported serving architecture determine those separate questions. Choose adapter-plus-base when the appointed software team can retain the exact common base and supports the required adapter selection. Choose complete merged variants when fixed task-specific deliveries suit the supported runtime and the retained controls meet the requirement. For example, if disabling adaptation to use the original base is required, receiving M1 and M2 alone does not establish that capability. The comparison turns on what the buyer needs to retain and change. A smaller transfer is valuable when the remaining dependencies are already covered. A complete per-task weight delivery is valuable when it resolves a dependency the buyer would otherwise have to supply. ## Compare what remains available after the initial handoff For an adapter arrangement, identify how the exact base remains available for a later deployment. A reference to a remote model is different from an agreed retained copy. If the buyer's environment cannot retrieve it when needed, possessing the adapters does not close that gap. Describe the supported runtime and library versions, artifact formats, model configuration, tokenizer and any other declared inference dependencies for either arrangement. PEFT's Quicktour illustrates separate tokenizer selection in its inference example. A complete weight set does not by itself describe the entire application environment. Then price a change to the first task. Does the supplier return a replacement adapter for the retained B1, or a replacement complete M1? If the base is to change, who establishes the newly supported pairing or merged variant? Keeping these responsibilities explicit makes the later supply comparable without assuming that every adapter can move to the next base revision. File possession alone establishes neither licensing rights nor an agreement to operate an inference service. Keep the applicable base and adapter permissions with their responsible owners, and state any continuing software supply or serving work separately. The appointed application team must also assess evidence for the intended task and workload; an artifact list reports supplied objects, not a live inference result. References: [H3](https://huggingface.co/docs/peft/quicktour) ## Connect a separate equipment-installation need to its physical scope If the same project also needs GPU equipment installed, use Leviathan's Complete Physical Deployment service and its project inquiry to discuss receiving, rack integration, agreed hardware QA, onsite installation and cabling. Bring a nonconfidential equipment list, quantities, destination, target window and the physical records required by the receiving team. Model conversion, application hosting and inference-service operation remain with the separately appointed software providers in this comparison. ## Primary sources - [R1 — Hu et al. — LoRA: Low-Rank Adaptation of Large Language Models, v2, section 4.1](https://arxiv.org/html/2106.09685v2). Checked 2026-10-20. Frozen base matrix and low-rank update dimensions. The 1024-by-1024, rank-8 example is original single-layer arithmetic; paper benchmark and resource ratios are not adopted. - [H1 — Hugging Face PEFT — PEFT checkpoint format](https://huggingface.co/docs/peft/developer_guides/checkpoint). Checked 2026-10-20. Adapter weight/configuration contents, base identity and revision fields, full-model storage and limitations of the merged-and-unloaded form. Applies to the documented PEFT arrangements rather than every adapter format. - [H2 — Hugging Face PEFT v0.21.0 — LoRA: Post-Training](https://huggingface.co/docs/peft/v0.21.0/package_reference/lora). Checked 2026-10-20. Distinguishes a standalone merged result from a merge retaining adapter machinery. No merge procedure, performance claim or configuration recommendation is adopted. - [H3 — Hugging Face PEFT — Quicktour: Inference and multiple adapters](https://huggingface.co/docs/peft/quicktour). Checked 2026-10-20. Separate tokenizer selection and adapter selection illustrate dependencies and capabilities the software supplier must specify. Example sizes, code and task outputs are not adopted. ## Related Leviathan resources - [Quantized weights and the complete inference-memory requirement](/articles/quantized-model-weights-versus-inference-memory) - [Training checkpoints and storage snapshots preserve different objects](/articles/gpu-training-checkpoint-versus-storage-snapshot) - [Local GPU software artifacts and their continuing dependencies](/articles/disconnected-gpu-deployment-artifacts-and-licenses) - [Complete physical infrastructure deployment](/services/turnkey-gpu-deployment) - [Discuss a separate physical GPU installation project](/contact?service=turnkey-gpu-deployment&from=%2Farticles%2Flora-adapter-versus-merged-model-delivery) ## Need the GPU equipment physically deployed? Share a nonconfidential equipment and site outline, the requested physical stages and the receiving team's handoff needs to discuss scope and availability. [Discuss the scope](/contact?service=turnkey-gpu-deployment&from=/articles/lora-adapter-versus-merged-model-delivery)