Hinode: batch inference on as many GPUs as you need
Put the dataset in one shared folder and start one job per part. Each job gets its own GPU. Label, caption, classify or score with vLLM, Ollama or your own script, billed by the minute.
Machines from $1.75/hr · Billed by the minute ·
Pauses itself when done
Sign up here. Start the jobs from your computer.
Started from workspace v6
$ for i in 0 1 2 3 4 5 6 7; do
hinode run --size 24x32 \
--name label-$i -d \
"python label.py --part $i --of 8"
done
8 jobs started
part 0
part 1
part 2
part 3
How it works
Set up once, then give every part its own GPU
Every machine starts from your workspace snapshot and reads the dataset from the same shared folder. Nothing is uploaded twice, and nothing waits in a queue.
Set up once · on the workspace
~ — zsh$ hinode apps install vllm
$ hinode env promote
snapshot v6 saved
Each part · on its own machine
label.py --part 3 --of 8Succeededlabel-3 · 2h 07m · $3.70
label-3 paused itself 15 min laterWhen the last part ends · your inbox
Hinode4:36 pmYour jobs finished: 8 of 8 succeeded
All eight parts exited 0 on label-0 to label-7. Each machine paused itself after its part. The results are in ~/shared/out.
Why Hinode
Not a queue, not a pod, not someone else's model
Three ways people run a dataset through a model today, and what changes here.
Instead of a cluster queue
No waiting for a slot
On a shared Slurm cluster, a job waits until a GPU frees up. Here each hinode run gets its own machine the moment you start it.
Instead of a rented pod
No setup on every machine
A pod starts empty, and the dataset goes up again each time. Here every machine starts from your snapshot, and the data is already in the shared folder.
Instead of a batch API
Your model, your code
A batch API runs the models it sells. Here it is any open model and your own script: vLLM, Ollama, PyTorch or anything pip installs.
Eight machines for an hour cost the same as one machine for eight hours. Splitting the dataset buys time, not a bigger bill.
Pricing
Cloud GPU pricing, by the minute
No plan and nothing up front. One bill at the end of the month, itemised by what ran.
WORKSPACE
$0.297/hr
A desktop on 1/8 of an NVIDIA L4, your files and your apps. Pauses itself when you leave.
- Disk
- $0.15 / GB-month written
- Shared folder
- $0.11 / GB-month
- Traffic out
- $0.10 / GB
MACHINE · 24 GB
from$1.75/hr
NVIDIA L4
Image generation, light fine-tuning, most ComfyUI work
MACHINE · 48 GB
Popularfrom$3.30/hr
NVIDIA L40S
Video models, bigger LoRAs, 30B-class inference
MACHINE · 96 GB
from$8.05/hr
RTX PRO 6000 Blackwell
70B-class models, long video, what fits nowhere else
What a month costs
≈ $112/month
87 h of workspace $25.84 + 26 h on 48x32 $85.80. Disks and the shared folder are extra.
These are the rates your invoice uses. The US and Europe cost the same.
FAQ
Batch inference on Hinode: questions
Any open model and any code that runs on Linux with an NVIDIA GPU. vLLM, Ollama and PyTorch install from the app catalog in one click, and pip, conda and apt work as usual. Install on the workspace, save a snapshot, and every machine starts with it.
Your script does. Pass it a part number, such as --part 3 --of 8, and let it take every eighth file. Each job writes its own output file to the shared folder, so no two jobs write to the same place.
No limit is set per account. If a region has no free card of the size you asked for, the machine waits and tries again, and you can pick another size meanwhile.
Each machine bills by the minute while it runs, from $1.75 an hour for 24 GB of VRAM. Eight machines for an hour cost the same as one machine for eight hours. When its job ends, a machine pauses itself after 15 idle minutes.
Only that part runs again. Its machine keeps everything installed, so you resume it and give it the same command. Make your script skip items that already have a result, and a rerun picks up where it stopped.
In the shared folder, which the workspace and every machine read. It costs $0.11 per GB-month for what is in it, and deleting a machine does not touch it.
It does the same job without the queue. Each hinode run gets its own machine right away, --after makes one job wait for another, and hinode job list shows how every part ended.
Yes. Connect Claude or ChatGPT to your account and ask in your own words. It quotes the price, attaches the machines, starts a job on each and tells you when the last one ends.
Put the dataset in once. Run it on as many GPUs as it needs.
Create a workspace, install your model, and start one job per part.
