Article

SageMaker AI full fine-tuning vs LoRA: how to choose

For teams customizing models, compare SageMaker’s supported recipes, training costs and continuation limits before choosing full fine-tuning or LoRA.

Editorial illustration for SageMaker AI full fine-tuning vs LoRA: how to choose: a geometric block represents a model release. Not documentary evidence.

AWS added full fine-tuning to SageMaker AI’s serverless model customization on 3 August 2026, giving teams adapting models for their business an alternative to LoRA adapters. A practical starting point is LoRA; consider full fine-tuning when substantial, high-quality task data and a demonstrated quality gap justify changing the whole model.

The choice is not simply which method scores higher. Full fine-tuning produces larger model files and cannot use SageMaker’s documented LoRA-only continuation workflow. Those constraints may decide the project before a benchmark does. AWS’s full fine-tuning documentation sets out these differences.

When is LoRA enough?

First check whether training is necessary. If clearer instructions solve the problem, improve the prompt. If the model needs access to company documents, consider retrieving that information at answer time. AWS’s customization guide distinguishes these approaches from teaching new behavior through training.

LoRA, short for low-rank adaptation, keeps the original model weights frozen and trains small adapters. Unmerged adapters still need the base model to run, but several task-specific adapters can share that base. This makes LoRA a useful starting point for a limited-budget comparison or a service with multiple specialized tasks.

Full fine-tuning, or FFT, updates all model weights and produces a full-weight model artifact. AWS recommends considering it for substantial domain changes with large, high-quality datasets, while warning that small datasets can cause the model to lose existing capabilities. There is no verified universal number of examples that makes FFT the right choice.

Do not choose FFT solely to avoid adapter overhead when serving responses: SageMaker also supports merging LoRA adapters into the base model. Whether to retain small, interchangeable adapters or deploy merged weights is a separate decision.

Will you need another training stage?

This is a concrete service constraint. SageMaker’s continuous-customization and resume-training workflows support LoRA only. Continuous customization starts from previous adapters with fresh training state; resume training restores an interrupted job’s checkpoint state.

If your plan is to teach the model from examples, then refine it using preferred versus rejected answers, the documented managed continuation route needs LoRA. That is a limitation of this SageMaker workflow, not a claim that full-weight models can never be trained further elsewhere.

Check the model, region and account before budgeting

Support checked on 19 November 2026 is specific to the model, learning technique and training type. A model’s presence in SageMaker does not mean every combination is available.

For proprietary data, also check evaluation routing. AWS says Bedrock-powered evaluation features may process data across regions within a geography. A training-region choice does not by itself settle where every evaluation step runs.

Prepare comparable data, not identical settings

For supervised fine-tuning (SFT), the model learns from examples of inputs and desired outputs. SageMaker accepts JSONL data—one JSON record per line—in conversation-message or prompt/completion form. Switching between LoRA and FFT for the same technique does not require a new dataset schema.

Keep the reviewed training data and held-out evaluation set fixed for a comparison, but use method-appropriate settings. The SFT parameter reference separates the two methods: FFT does not use LoRA rank, alpha or dropout settings. Do not assume the same learning rate will suit both.

Inspect duplicates, labels and task relevance before paying for either method. The SageMaker dataset-quality guide explains how to review an existing dataset; those preparation tools have their own regional availability and charges.

What does a LoRA-versus-FFT comparison cost?

AWS bills SFT and direct preference optimization by training tokens: dataset tokens multiplied by epochs, or passes through the data. Reinforcement-learning customization instead uses a model-dependent hourly rate, billed per second for partial hours. Do not apply SFT token rates to a reinforcement-learning job.

For Qwen3 4B SFT in US East (Northern Virginia), the official pricing table checked on 19 November lists US$0.50 per million training tokens for LORA and US$0.75 for FULLRANK, the full fine-tuning option.

Illustrative calculation, not a training result: assume a dataset containing 10 million billable tokens, three epochs, and one completed job per method at those list rates. Each job processes 30 million training tokens.

  • LoRA: 30 × $0.50 = $15.00.

  • Full fine-tuning: 30 × $0.75 = $22.50.

  • One job of each type: $37.50 in training charges.

FFT’s training charge is 50% higher in this example, a $7.50 difference. This assumes no retries, parameter sweeps, discounts or tax adjustments. Ten million tokens is an arithmetic input, not a recommended dataset size. Count tokens in the actual data; a record count is not a token count.

Evaluation is charged separately for input and output tokens, with additional Bedrock usage if a model acts as judge. The subtotal also excludes data preparation, storage, deployment, inference and staff time. Other models, regions and techniques need their own price checks.

What AWS’s coding results do—and do not—show

AWS’s comparison dated 18 November is a new example, not a new FFT launch. It describes Qwen3-4B coding experiments with 99,000 training examples and separate validation and test sets of 5,500 each.

AWS reports a 90.0% model-judged score for one FFT setting. Elsewhere in the same article it reports LoRA at 81.5% versus FFT at 79.5%, and mathematics performance falling below the base model after both methods. These differently described settings should not be combined into one leaderboard.

The distinction between metrics matters. AWS’s PRIMECODE scores measure resemblance to reference text, while Amazon Nova Pro supplies a model’s judgment of correctness. Neither is an executed-code pass rate. The results support running a comparison for your task, not assuming FFT will win or that 90% of generated programs run correctly.

Independent research published in TMLR found stronger target-domain learning from full fine-tuning but better preservation of other tasks from standard low-rank LoRA in its experiments. That is context, not a replication of the SageMaker comparison or a guarantee that LoRA preserves every capability.

What would justify choosing full fine-tuning?

Set the acceptance rule before funding the comparison: the task improvement required, the capabilities that must not regress, and the total experiment budget. Keep the untouched base model as a reference. For code generation, include execution-based checks appropriate to the task rather than relying only on text similarity or a model judge.

The guide to fine-tuning budgets, baselines and model review offers a complementary framework for bounding an experiment and reviewing its output.

Choose FFT only if it clears your task-quality threshold, preserves the other capabilities you need, fits the full operating budget and leaves a workable path for future training. If LoRA already meets the requirement, a larger full-weight artifact and another experiment need a specific business justification.