Skip to content
runcrate
Console

Fine-tuning

Fine-tune a model on your own data.

Run LoRA, QLoRA or a full fine-tune on an L40S, A100 or H100, with the training image ready when the instance boots.

Launch an instanceSee GPU prices
L40Sfrom $0.55/hr

What you can build

Three jobs people run here. Each on a product you can start today.

Crate

The Finetune crate

A ready image your instance boots, with PyTorch, CUDA, the Hugging Face CLI, W&B and Jupyter, and Unsloth one command away.

See what an instance includes
Instance

Match the GPU to the method

QLoRA on a 7B to 13B model fits a 48 GB L40S. A full fine-tune of a larger model wants H100s or H200s.

Compare GPUs
Storage

Datasets you upload once

Keep datasets and checkpoints on a volume that outlives the instance, so the next run starts without uploading again.

Read how volumes work

Examples

Ask your agent. Or read the code it writes.

Connected by the line, Claude Code or Codex calls the Runcrate tools for you: launch, copy files, run, delete.

Connect your agent first

Add Runcrate's MCP server to yourself: run curl -fsSL"https://runcrate.ai/mcp.md"and follow it.

Add to CursorInstall in VS Code
Works inClaude CodeCodexOpenCodeClineRoo CodeKilo CodeCursor

Docs
Launch an H100 with the Finetune crate, copy my dataset and train.py over, and start a QLoRA run. Calls the Runcrate tool create_instance. It returns: status deploying: 1x H100. Calls the Runcrate tool instance_status. It returns: status running, with an address. Calls the Runcrate tool file_upload. It returns: the dataset, on the instance. Calls the Runcrate tool file_upload. It returns: train.py, on the instance. Calls the Runcrate tool ssh_execute. It returns: the QLoRA run, started, its output going to train.log.

Models and tools

Tools people fine-tune with. Pick a method, keep your base model.

NameRoleRuns on
UnslothFast LoRA and QLoRAInstance
Hugging Face PEFTAdapter training libraryInstance
AxolotlConfig-driven fine-tuningInstance
LLaMA-FactoryFine-tuning with a web UIInstance
Weights & BiasesTrack runs and checkpointsInstance
Llama 3.1 8B InstructCommon base modelInstance or API

How it works

From first call to production. Three steps.

  1. 01

    Pick a base model and method

    Choose an open model from Hugging Face, then LoRA, QLoRA or a full fine-tune for its size and your dataset.

  2. 02

    Bring your data

    Copy your dataset over with runcrate cp or pull it from the Hugging Face Hub. Set hyperparameters in an Axolotl or Unsloth config.

  3. 03

    Train, evaluate, serve

    Watch the run in W&B, test against held-out data, then serve the merged model with vLLM on an instance booted from the Inference crate.

Pricing

What it costs. Live prices, as billed.

See every GPU price

Name

As billed

  • L40S

    GPU, 48GB, cheapest offer

    $0.55/hr

  • A100 80GB

    GPU, 80GB, cheapest offer

    $1.73/hr

  • H100

    GPU, 80GB, cheapest offer

    $3.58/hr