onrup

Granite · IBM

Fine-tuning Granite 3.0 8B

What does it take to fine-tune Granite 3.0 8B?

Granite 3.0 8B needs 18 GB for half-precision LoRA training and 14 GB in four-bit, and 16 GB to serve. It supports supervised fine-tuning, preference tuning (dpo), under the Apache 2.0 licence. The cheapest qualifying class is RTX 4000 Ada at $0.09 per GPU-hour.

The 8B in the Granite 3.0 line, with the same enterprise orientation as the 2B and enough capacity to hold real domain knowledge. Apache-2.0, documented training data, and behaviour tuned to decline rather than improvise when it does not know.

What to know before choosing it

The same enterprise orientation as the 2B with enough capacity to hold real domain knowledge. Apache-2.0, documented training data, and a disposition toward refusing rather than guessing — the combination is what makes it a common choice for internal question answering over proprietary documents.

Against Qwen3 8B or Llama 3.1 8B on general benchmarks it will usually lose. On a retrieval-grounded task with a refusal requirement it frequently wins, and that is the comparison worth running through the gate.

Specification

Granite 3.0 8B specification
Parameters8B
ArchitectureDense
Size tierMid
LoRA training memory18 GBHalf precision, frozen base
QLoRA training memory14 GBFour-bit base, higher-precision adapter
Serving memory16 GBHalf precision, before attention cache
ObjectivesSFT, DPO
LicenceApache 2.0
Repositoryibm-granite/granite-3.0-8b-instruct

What it costs to train

Every class with enough memory for four-bit training, cheapest first. Total cost is the rate multiplied by wall time, so the cheapest rate is not always the cheapest run — a faster class that finishes sooner frequently wins.

GPU classVRAMTraining / hrServing / hrFits
RTX 4000 Ada20 GB$0.09$0.11LoRA and QLoRA
L424 GB$0.17$0.20LoRA and QLoRA
RTX 309024 GB$0.22$0.26LoRA and QLoRA
RTX 409024 GB$0.38$0.43LoRA and QLoRA
A4048 GB$0.42$0.48LoRA and QLoRA
RTX 6000 Ada48 GB$0.61$0.71LoRA and QLoRA
A600048 GB$0.65$0.75LoRA and QLoRA
L40S48 GB$0.78$0.90LoRA and QLoRA
A100 40 GB40 GB$1.17$1.35LoRA and QLoRA
A100 80 GB80 GB$1.68$1.94LoRA and QLoRA
H100 80 GB80 GB$2.59$2.99LoRA and QLoRA
H200141 GB$4.55$5.25LoRA and QLoRA

Serving fits on RTX 4000 Ada at $0.11 per hour — before the attention cache, which grows with context length and concurrency.

Good starting point for

Preference tuning on this model

Preference tuning holds a frozen reference copy of the model alongside the one being trained, so budget roughly 36 GB rather than 18 GB. That is the thing that catches people out — supervised training on this model fits on a class that preference tuning will overflow.

When preference tuning beats supervised fine-tuning →

Other sizes in this family

ModelParamsQLoRAObjectives
Granite 3.0 2B2B5 GBSFT, DPO
Granite 4.1 8B8B14 GBSFT, DPO

Comparable sizes elsewhere

Frequently asked questions

How much VRAM does it take to fine-tune Granite 3.0 8B?

18 GB for half-precision LoRA and 14 GB for four-bit QLoRA. Serving needs 16 GB. Preference tuning roughly doubles the training figure, because a frozen reference model is held alongside the one being trained.

What is the cheapest way to fine-tune Granite 3.0 8B?

Four-bit QLoRA on RTX 4000 Ada at $0.09 per GPU-hour is the cheapest class that meets the 14 GB threshold. Note that quantised training is slower per step, so a faster class sometimes costs less over the whole run.

Can I download the weights after fine-tuning Granite 3.0 8B?

Yes. Every finished run exposes its trained weights for download, and publishing to a model hub is a single call with a generated model card recording the base model and version the adapter applies to.

What licence does Granite 3.0 8B carry?

Apache 2.0. The licence follows the fine-tune — a derivative inherits the base model’s terms, and those terms pass to anyone you give the model to.

Last verified 6 August 2026. Memory thresholds are the platform's own admission limits.

Fine-tune Granite 3.0 8B

Upload a dataset, forecast the run, and see the cost before any compute is leased.