Bootstrapped SaaS startup: New limits on fine tuning on Helix free plan
On trying to make a SaaS service pay for itself while bootstrapping
Hey users & subscribers,
We’ve had an big uptick in number of fine-tunes on the SaaS platform in recent weeks, which is great, but it also costs money. We’re using together.ai for the qapair generation with Mixtral, and we just got a surprise $1K bill.
What’s more, the GPUs we run the finetunes on cost money too, and our runpod bill is about $2-3K per month. I’m currently paying both of these bills out of my personal savings (plus the electricity to run the two GPUs I run in my home data center!).
When long-running finetunes eat up all our GPU resources, we can’t serve inference requests, and increasingly we’ve had all our GPUs allocated to fine-tuning. I’m happy users are finding the fine tuning service interesting and hopefully useful enough to kick off those runs!
But tl;dr - We’re a bootstrapped startup, and so in order to keep the SaaS running we’ve had to limit the size of documents you can fine-tune for free. Free-tier finetunes are now limited to 5 chunks, about one page of text. To finetune larger datasets (up to 1000 chunks), please upgrade to the premium tier which is just $20/month. You can upgrade at https://app.tryhelix.ai/account. To train above 1000 chunks (~200 pages), please get in touch.
Even if you just like what we’re doing, please subscribe to support us!
Go on, click that sweet subscribe button 😅
New stuff we’ve done recently:
Tools launch & demo at Ollama meetup in Paris (8 minutes)
GPTScript in microVMs in Helix (including in the SaaS):
New “full platform” vision articlated on updated homepage:
As ever, any complaints or comments come find us in Discord please :-)
Cheers,
Luke
CEO, HelixML