There is a persistent myth that working with large language models means renting time on someone else's farm of expensive accelerators. For one specific, very useful job — fine-tuning a small model on your own data — that myth is wrong. A 7B or 13B model can be fine-tuned on a single consumer graphics card sitting under your desk in Lagos, Abuja or Port Harcourt, provided you choose the right method and the right hardware. This guide walks through exactly what that hardware looks like, what it costs in Naira, and where the real limits sit.
First, a clarification that saves a lot of wasted money. Fine-tuning adapts an existing open model to your tone, domain or output format — it is not training a model from scratch. Training from scratch genuinely does need a cluster, and you should not attempt it locally. If you want the deeper theory behind the efficient methods we lean on here, read our LoRA and QLoRA training hardware deep-dive, and for the business case of keeping this work in-house see our guide to fine-tuning LLMs on in-house hardware.
What fine-tuning actually changes
Fine-tuning takes a capable base model — say a 7B parameter open model — and nudges its behaviour using a few hundred or few thousand examples of the input-output pairs you care about. You are not teaching it language from nothing; you are teaching it your house style, your product catalogue's vocabulary, your support tone, or a strict output format. The base model already knows how to write. You are specialising it.
This distinction matters because it sets the hardware bar. From-scratch training is memory-brutal and compute-brutal. Adapting an existing small model, done the efficient way, is something a single good GPU handles comfortably over a few hours.
Why full fine-tuning is the wrong path locally
The naive approach — updating every weight in the model, called full fine-tuning — is punishing on memory. You do not just need room for the model's weights. You also need room for the gradients and for the optimiser states, which together often consume several times the size of the model itself. A 7B model that occupies roughly 14GB at half precision can balloon to tens of gigabytes of VRAM once gradients and optimiser states are added. That is well beyond any consumer card.
The result is simple: almost nobody full fine-tunes locally. It is reserved for organisations with multi-card server hardware and a specific reason to update every weight. For the rest of us, there is a far smarter route.
The practical path: LoRA and QLoRA
Parameter-efficient fine-tuning is the technique that makes local work possible. Instead of updating the whole model, you freeze the original weights and train small adapter layers on top. LoRA does this and slashes the memory needed for gradients and optimiser states, because you are only optimising a tiny fraction of the parameters.
QLoRA goes further. It loads the frozen base model in 4-bit precision, shrinking its memory footprint dramatically, then trains LoRA adapters on top of that quantised base. The combination is the headline of this whole article: with QLoRA you can fine-tune a 7B model on a single 24GB consumer card. A used RTX 3090 or an RTX 4090 — both 24GB — is the local fine-tuning sweet spot, and the second-hand market makes the 3090 especially attractive. Our breakdown of the used GPU market, RTX 30 versus 40 series, covers what to pay and what to avoid.
Tooling has caught up too. Frameworks like axolotl, Unsloth and the PEFT library wrap this process so you are writing a short configuration file rather than low-level training code. The barrier now is hardware and a clean dataset, not deep machine-learning engineering.
The VRAM ladder for QLoRA
VRAM is the single number that decides what you can fine-tune. Here is a rough ladder of what fits via QLoRA, from modest to serious. For a fuller treatment of why VRAM dominates, see GPU VRAM for training models and our general guide to what GPU VRAM you need in 2026.
- 12GB — tiny models (around 1B to 3B) fine-tune comfortably; a 7B is a tight squeeze with aggressive settings and short sequences, and often frustrating.
- 16GB — a 7B fits with QLoRA at modest sequence lengths and batch sizes. Workable, but you will be careful with settings.
- 24GB — the sweet spot. A 7B fine-tunes with room to spare; a 13B is achievable with careful sequence length and batch settings. This is where most serious local work happens.
- 48GB — a single large card or two 24GB cards together. Comfortable 13B work and a path towards larger models. Beyond this, you are looking at multi-GPU rigs or the cloud.
If your ambitions outgrow a single 24GB card — bigger models, or several jobs in parallel — a two-card build is the next step. We document one end to end in our dual-GPU AI training rig build.
The forgotten half: the dataset-prep PC
Newcomers fixate on the GPU and forget that a large slice of any fine-tuning project is preparing the dataset. Collecting examples, cleaning them, de-duplicating, formatting them into the input-output structure the trainer expects, and validating that they are correct — this is real work, and it is mostly CPU, RAM and storage work, not GPU work.
The good news is that an ordinary good workstation handles it: a solid multi-core processor, 32GB or more of system RAM, and a fast NVMe drive for shuffling data files around. You do this part long before the GPU ever spins up. If you are speccing a machine that does both the prep and the training, our AI training workstation for business guide pulls the whole build together.
Rough Naira tiers
Prices move, so treat these as bands rather than quotes. They reflect a complete working machine, not just the card.
- Entry single-24GB rig — built around a used RTX 3090, a capable CPU, 32GB to 64GB of RAM and fast storage. This is the realistic starting point for serious local 7B and 13B fine-tuning, typically in the low-to-mid millions of Naira depending on the rest of the build and current used-card prices.
- Dual-GPU rig — two 24GB cards, a workstation-class platform with enough PCIe lanes, a larger power supply and serious cooling. This pushes well up into the multi-millions and is justified when single-card jobs are too slow or too small for your needs.
When to rent the cloud instead
Local hardware is not always the right answer, and it is worth being honest about that.
- One-off jobs. If you need to fine-tune once and never again, renting a cloud GPU for a few hours is cheaper than buying a card.
- Bigger models. Once you are past what 24GB or 48GB can hold, the economics of buying many cards stop making sense for occasional work.
- No stable round-the-clock power. This is the Nigerian reality check. If you cannot guarantee power for the length of a run, the cloud removes that risk entirely.
For a structured comparison of owning versus renting, including the running-cost maths, read on-premise AI compute versus cloud for business.
NEPA, UPS and the mid-run crash
Fine-tuning runs for hours, not seconds. That single fact makes power stability non-negotiable here. If NEPA drops mid-run and your machine cuts out, you do not lose a few seconds — you can lose the entire run, every hour of GPU time spent up to that point, unless you were diligently saving checkpoints. A proper UPS sized to hold the rig through brief outages, and ideally an inverter or generator handover for longer ones, is not a luxury for this workload. It is the difference between a finished adapter and a wasted afternoon.
Plan for stable power the same way you plan for VRAM. A 24GB card you cannot keep powered for four uninterrupted hours is a 24GB card that cannot finish a job.
Frequently Asked Questions
Can I really fine-tune a 7B model on one consumer GPU? Yes — with QLoRA on a 24GB card such as a used RTX 3090 or an RTX 4090, a 7B fine-tune is comfortable, and a 16GB card can manage it with careful settings. Full fine-tuning is a different story and is not realistic on consumer hardware.
How long does a fine-tune take? It depends heavily on the model size, dataset size, sequence length and your card, so any single number would be misleading. The honest answer is hours rather than minutes for a typical small-model job — which is exactly why stable power and checkpointing matter so much.
Do I need a powerful PC just to prepare the dataset? No. Dataset preparation is CPU, RAM and storage work that runs fine on an ordinary good workstation — a decent multi-core CPU, 32GB or more of RAM, and a fast NVMe drive. The GPU only earns its keep during the training run itself.
The One Thing to Remember
If you take away a single idea, make it this: you do not need to update the whole model, and you do not need a datacentre. QLoRA on a single 24GB consumer card lets you adapt a 7B — and with care a 13B — to your own data, at home, in Naira you can actually budget for. The technique, not the cluster, is what unlocks local fine-tuning.
Ready to build a machine that can do this? Spec a single-card or dual-GPU fine-tuning rig with our configurator, or contact us to talk through your model, your dataset and your power situation before you buy.