That’s with us now.

Step 1 of 4
Draft saved

Tell us what you’re building.

Plain language is fine — better, in fact. We read every proposal ourselves. How this works →

Not ready to fill this in yet? Create an account and come back to it whenever you like — you can start a project from your dashboard at any time.

Give your project a title.
This is the line we read first, and the one that appears on your project page.
0 / 140
One sentence is enough — but we need it.
Tell us a little about the work.
Be concrete. “Cuts triage time at 3 clinics from 20 minutes to 5” tells us more than “improves healthcare access”.
Who benefits, and how would you know it worked?
Pick the closest one.
Who benefits, and what do they get? Be specific about both.
Tell us who this is for, and what they get.
What stops this happening without us?
A sentence or two is plenty.

About you

We check who the compute is going to before we allocate it.

Pick the closest one.
Where are you based?
A departmental page, GitHub, LinkedIn, personal site, or a published paper. We ask everyone — it’s how we tell our infrastructure partners who the compute is going to.
We need one link we can check.
Briefly — your training, relevant experience, and any prior work in this area. A few lines is enough.
A few lines is enough.
Where should we send your link back? *
Required — it is how you finish this later, and the only way we can reply. This does not create an account.
We need an address we can reach you at.

What will it take to run?

If you already know your numbers, give them. If you don’t, describe the work and we will size it with you.

I have an estimate
You know roughly the hours and hardware.
I’m not sure yet
Describe the workload instead.
Across all GPUs, not per GPU.
Give a number, or choose “I’m not sure yet”.
Most projects run on one.
Enter the hours and we’ll show how long that runs for.
Name the specific model, not the family. “Qwen2.5-7B-Instruct”, not “Qwen”. If you’re comparing several, list them.
What you want to train or run, on how much data, and how many times. We will turn that into an estimate with you.
A few sentences is plenty.
Name the specific model, not the family. “Qwen2.5-7B-Instruct”, not “Qwen”. If you haven’t chosen yet, say so and list the candidates.
Name the model, or the ones you are choosing between.
Please answer — it changes how we set the machine up.

How ready is it to run?

This tells us how much setup help you need — not whether you qualify.

Pick the closest one.
When could you begin?
Is your code in a container?
Docker or similar. If not, we help you package it.
Can training resume from a checkpoint?
Matters if a run is interrupted.
Do you have a reliable connection to upload data?
We can arrange another route if not.

What will exist at the end?

Compute is donated on the understanding that something open comes back.

Add a line for each one. Models, datasets, papers, tools.
OutputExpected by
Add at least one output.
This list becomes your agreed outputs. We will not ask you to confirm it again.
Will the outputs be public?
If something must stay closed, tell us why below.
Licence
How others may use what you release.
This doesn’t affect eligibility.
Please pick one.
Including anything still pending. Some providers appear in more than one programme, so this helps us avoid duplicating an award you’ve already received.

And finally, about you.

Your affiliation and background are on step 1 — this is just where we write to.

We need a name.
What is your role?
Which country?
From step 1 — change it here if it was wrong.
We need an address we can reach you at.

Does this look right?

Nothing is sent until you press the button at the bottom.

What happens after you send this
A person reads it — usually within a week.
If we take it forward, we set up your account and send you the login. You do not need to create one.
If we cannot, we say so plainly and tell you why.