Skip to content
runcrate
Console

Model training

Distributed model training from one H100 to a cluster.

Start a training run on an H100 or B200 instance, keep checkpoints on a volume. Launch 8 or more GPUs yourself, or reserve a multi-node cluster.

Launch an instanceReserve a cluster
H100from $3.58/hr

What you can build

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

Instance

Multi-GPU runs on one instance

Ask for two or more GPUs on one instance and run DeepSpeed ZeRO or FSDP across them. You pay for the time it runs, or prepay a machine for 1 to 12 months at a discount for paying ahead, billed by Runcrate.

Compare training GPUs
Storage

Checkpoints on a volume

Write checkpoints to a volume that outlives the instance, so a stopped run resumes from its last save on a fresh instance.

Read how volumes work
Cluster

Reserved multi-node clusters

A multi-node cluster for a month or longer: we verify your company and your need, then take your exact spec to GPU owners that fit, and you see their offers side by side.

Reserve a cluster

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 two H100s with the Finetune crate, copy train.py over and start it with torchrun on both GPUs. Calls the Runcrate tool create_instance. It returns: status deploying: 2x H100. Calls the Runcrate tool instance_status. It returns: status running, with an address. Calls the Runcrate tool file_upload. It returns: train.py, on the instance. Calls the Runcrate tool ssh_execute. It returns: the run, started on both GPUs, its output going to train.log.

Models and tools

Frameworks people train with. All of them run on your instance.

NameRoleRuns on
PyTorchTraining framework, in the crateInstance
DeepSpeedZeRO sharding across GPUsInstance
FSDPFully sharded data parallelInstance
Megatron-LMTensor and pipeline parallelismInstance
Weights & BiasesLoss curves and run trackingInstance

How it works

From first call to production. Three steps.

  1. 01

    Pick the GPU

    Choose H100, H200 or B200 by name and how many you need. The Finetune crate boots with PyTorch, CUDA and W&B in place.

  2. 02

    Start the run

    Copy your script and data over, start torchrun or DeepSpeed over SSH, and watch loss in W&B and GPU load in nvtop.

  3. 03

    Keep the weights

    Save checkpoints and final weights to a volume or push them to Hugging Face, then delete the instance to stop the bill.

Pricing

What it costs. Live prices, as billed.

See every GPU price

Name

As billed

  • H100

    GPU, 80GB, cheapest offer

    $3.58/hr

  • H200

    GPU, 141GB, cheapest offer

    $5.13/hr

  • B200

    GPU, 180GB, cheapest offer

    $10.69/hr