onrup

Llama · Meta

Fine-tuning Llama 4 Scout 109B

What does it take to fine-tune Llama 4 Scout 109B?

Llama 4 Scout 109B needs 220 GB for half-precision LoRA training and 80 GB in four-bit, and 80 GB to serve. It supports supervised fine-tuning, under the Llama 4 Community licence. The cheapest qualifying class is A100 80 GB at $1.68 per GPU-hour.

Sixteen experts, seventeen billion active per token, and natively multimodal. Note the asymmetry in the numbers: serving fits in 80GB but half-precision training needs 220GB, so the practical path is four-bit training on an 80GB class and serving on the same. Experimental in this catalogue.

What to know before choosing it

Read the memory figures carefully, because they are not in the order anyone expects: serving fits in 80GB while half-precision training needs 220GB. The practical path is four-bit training and half-precision serving on the same class, which is an unusual shape and worth planning for.

Native multimodality is the reason to be here at all. If your task is text-only, a dense model a fraction of this size will be cheaper to run and easier to fine-tune, and the sixteen-expert routing will not pay for itself.

Specification

Llama 4 Scout 109B specification
Parameters109B17B active per token
ArchitectureMixture of experts
Size tierExtra large
LoRA training memory220 GBHalf precision, frozen base
QLoRA training memory80 GBFour-bit base, higher-precision adapter
Serving memory80 GBHalf precision, before attention cache
ObjectivesSFT
LicenceLlama 4 Community
Repositorymeta-llama/Llama-4-Scout-17B-16E

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
A100 80 GB80 GB$1.68$1.94QLoRA only
H100 80 GB80 GB$2.59$2.99QLoRA only
H200141 GB$4.55$5.25QLoRA only

Serving fits on A100 80 GB at $1.94 per hour — before the attention cache, which grows with context length and concurrency.

Good starting point for

This is a mixture-of-experts model

109B total parameters with 17B active per token. Memory scales with the total; inference compute scales with the active count. That makes it attractive when memory is cheaper than compute for your workload, and unattractive when the reverse holds. Quoting only one of the two numbers is how these models get misrepresented in both directions.

What mixture of experts means →

Other sizes in this family

ModelParamsQLoRAObjectives
Llama 3.2 1B1B3 GBSFT, DPO
Llama 3.2 3B3B6 GBSFT, DPO
Llama 3.1 8B8B14 GBSFT, DPO, GRPO
Llama 3.3 70B70B48 GBSFT

Comparable sizes elsewhere

Frequently asked questions

How much VRAM does it take to fine-tune Llama 4 Scout 109B?

220 GB for half-precision LoRA and 80 GB for four-bit QLoRA. Serving needs 80 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 Llama 4 Scout 109B?

Four-bit QLoRA on A100 80 GB at $1.68 per GPU-hour is the cheapest class that meets the 80 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 Llama 4 Scout 109B?

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 Llama 4 Scout 109B carry?

Llama 4 Community. 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 Llama 4 Scout 109B

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