Skip to content
runcrate
Console

Edge and embedded AI

Train big, ship small from cloud GPU to device.

Train on an H100, then convert to ONNX, compile with TensorRT and quantize to INT8 on the same instance before the model ships to a Jetson or phone.

Launch an instanceSee GPU prices
L40Sfrom $1.26/hr

What you can build

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

Instance

Train and compile in one place

Train in PyTorch, export to ONNX and build TensorRT engines on the same GPU, then profile latency at FP16 and INT8.

Compare GPUs
Crate

Start from the Finetune crate

The Finetune crate is a ready image your instance boots, with PyTorch and CUDA in place. Add TensorRT, ONNX Runtime or OpenVINO on top.

See what an instance includes
Storage

Keep every build

Store checkpoints and exported models on a volume, so a retrain next month starts from the last good build.

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 L40S with the Finetune crate, run export.py to build the ONNX model, and copy model.onnx back to me. Calls the Runcrate tool create_instance. It returns: status deploying: 1x L40S. Calls the Runcrate tool instance_status. It returns: status running, with an address. Calls the Runcrate tool file_upload. It returns: export.py, on the instance. Calls the Runcrate tool ssh_execute. It returns: exit code 0: model.onnx, written. Calls the Runcrate tool file_download. It returns: model.onnx, in your folder.

Models and tools

The optimization toolchain. All of it runs on your instance.

NameRoleRuns on
TensorRTCompile engines for NVIDIA devicesInstance
ONNX RuntimeCross-platform model exportInstance
AutoAWQINT4 weight quantizationInstance
OpenVINOIntel CPU and edge targetsInstance
Core ML ToolsConvert models for iOSInstance

How it works

From first call to production. Three steps.

  1. 01

    Train in the cloud

    Train on an H100 or A100 with PyTorch, any architecture: vision, language or multimodal. Save checkpoints to a volume.

  2. 02

    Optimize and quantize

    Convert to ONNX, compile with TensorRT and quantize to INT8 or INT4. Calibrate on representative data and compare accuracy against latency.

  3. 03

    Export to the device

    Copy the optimized model back with runcrate cp and ship it to Jetson, Core ML or TFLite. Come back to retrain when you need to.

Pricing

What it costs. Live prices, as billed.

See every GPU price

Name

As billed

  • L40S

    GPU, 48GB, cheapest offer

    $1.26/hr

  • RTX 4090

    GPU, 24GB, cheapest offer

    $0.86/hr

  • A100 80GB

    GPU, 80GB, cheapest offer

    $1.43/hr