One command to get an H100. Your folder syncs both ways. When you're done, deploy as a production endpoint.
H100/A100
GPU fleet
<30s
Provisioning
4 clouds
Capacity routing
Per-second
Billing
$ cassian up --gpu h100
Provisioning... ready (28s)
$ cassian ssh
root@cassian-h100-01:/workspace$ python train.py
[Epoch 12/12] loss: 0.0023 | 14m 32s
root@cassian-h100-01:/workspace$ exit
$ cassian deploy
✓ https://api.cassian.cloud/v1/my-model
No quota tickets. No SSH sprawl. No lost work. No paying while nothing runs.
One command gets you an H100 in under 30 seconds. No quota requests, no waiting days for approval.
One working tree. Your local folder stays in sync with the remote GPU. No SCP, no manual uploads.
Mac, Windows, Linux. Edit locally, run on a GPU. We handle SSH, sync, and environment setup.
Spot instance goes down, volume reattaches on a new machine. Your work survives across sessions.
Picks the cheapest available GPU across CoreWeave, GCP, AWS, and Azure. You never file a quota ticket.
Training done? Run one command. HTTPS endpoint with TLS, autoscaling, and custom domains.
One YAML file. GPU type, CUDA version, disk, dependencies, what to sync. Replaces Dockerfiles, shell scripts, and setup docs.
name: llama-finetune
gpu: h100
cuda: "12.4"
disk: 500GB
python: "3.11"
sync:
local: ./
remote: /workspace
exclude: [".git", "data/raw"]
pip:
- torch>=2.3
- transformers
- accelerate
- wandb
Your folder is on an H100 in 30 seconds. Files sync both ways. Use SSH, VS Code, or Cursor to edit and run.
$ cassian up
Provisioning... ready (28s)
Sync: ./ ↔ /workspace
$ cassian ssh
root@cassian-h100-01:/workspace$ python train.py
[Epoch 1/12] loss: 2.341
[Epoch 2/12] loss: 1.847
...
[Epoch 12/12] loss: 0.003 ✓
Training done. One more command turns your model into a production inference endpoint with TLS and autoscaling.
$ cassian deploy --domain api.mycompany.com
Deploying...
✓ Live
https://api.mycompany.com/v1/llama-ft
$ curl https://api.mycompany.com/v1/llama-ft \
-d '{"prompt": "Explain transformers"}'
{"response": "Transformers are a neural..."}
We're working with teams who are tired of pods that crash, uploads that fail, and credits that disappear.
Talk to us