“don’t train your own model” is common ai advice. it's wrong. your token bill's the proof. today, we’re excited to launch castform into open preview. castform is the easiest way for you to train your own model, on your own data. open-weights models are performant and much cheaper. when... trained on your task & proprietary data, they beat closed models. the thing standing between you and that was weeks of plumbing & years of ml expertise. with castform, model training is as simple as prompt engineering. castform bring your agent traces or raw corpora. castform turns it into training data, picks the right algorithmic recipes, manages gpus, and gives you an ide to watch and chat with your model as it learns. see what you can build with castform👇show more

1/4 train your first model today. new users get $50 in free credits 🙂

2/4 you can train a model on your production agent traces so the model can learn from mistakes it made in product with castform.

3/4 you can train an agent to do accurate and fast citation-grounded search over documentation

4/4 you can train a model to red-team & find bugs with castform

train your first model today at

Me right now…. Congrats!!! Can’t wait to try it soon

🤣 you can try it at -> $50 in free credits for new users :) pls give us feedback!!

@sparab22 Can’t wait to try! Is it suitable for non technical peeps?

@sparab22 give it a go! exporting the starter template and letting claude code go wild has worked surprisingly well - let me know if you need help with setup :)

Love this Girish. We are building infra where you can use Prediction markets to do continuous RL workflows . We call this RLEF - Reinforcement Learning through Economic Feedback. Have some datasets and would love to train own model! @ChronisKod @HotTubLee

@ChronisKod @HotTubLee yes would love to! if you can drop me a note at [email protected] we can get on a call

@ChronisKod @HotTubLee 🙏

@berman66 so good!

@berman66 thanks brett! big fan of your launch video commentary :)

@berman66 yall won

@berman66 high praise - thanks so much!!

This is SICK

was great working with y'all on this!

54 times cheaper than a hundred billion dollars still seems like a considerable barrier to entry. But seriously, anything that makes model training more accessible is important. I've spent enough of my own time doing it to know the potential value of such services.

the 54 times cheaper is inference cost not training cost :) small models are surprisingly cheap to post-train, on the order of 100s of $$. and yes, our goal here's to make this as accessible as possible

we need more videos with castie!!!

yes more content with castie incoming - goal's the hopefully use castie as a way to better communication post-training research with devs :)

this is huge! why use a generalized model, when you can train your own with ease!

yup exactly! we make training easy so it's a no-brainer

Please make Castie merch!!!

@thitrangmam 👀👀

CASTIEEEEEE IS SO CUTE

yessss 100%

wow, this is huge

was great working with y'all on this!

Are you sure the Nintendo people can't come after you for the pokemon ref?

hahhaha

can these models perform complex tasks like image segmentation from limited data? what kind and how much data would one need for a self-trained model to be able to compete with the foundational ones?

@vatslgoswami currently, we support text models only -> multimodal coming by end of the summer :) on data, depends on the task but have seen pretty good results from 100s to 1000s of samples

this is really cool stuff. can't wait to play around with it! we needed image segmentation on food dishes. i'm wondering what amt of data something like that would need. on a side note, the blog mentions something about tools. you're telling me i can train my model to be more efficient with tools, etc.?

@vatslgoswami ahh my intuition for something segmentation would be mid 100s but could be wrong. and yes you can train models to be much more efficient with tool calls/usage

that is so doable!! maybe we will be on your multimodal customer stories sometime haha

small oss models!! >>

yes 100% for specialized tasks!

Congrats on the launch! 🚀

thanks guys!!

yo what the heck, animation go crazyyyy

yesss

this is awesome😍

thanks denise! goal's to make model training as simple as prompt engineeing :)

congrats! which open weight models do you currently support?

thanks joshua!! currently qwen, gemma's coming very shortly :)

Cute launch video! 🐙 Any details on rough pricing?

new users get $50 in free credits & pricing is token-based (and based on the size of the model)

What would be your recommendations for using this w/ trading?

ooh good question, not something we've looked into actually but i wonder if you can train a model to be good at forecasting -> e.g. given newspaper headlines, twitter discourse, economist speeches etc. predict if the fed will hike/cut rates

And yeah I was thinking about analyzing it just the way you said. BTW, can we choose the model we want to train? Or we'd be training your specific model?

we currently have support for qwen -> gemma's gonna be out shortly :)

Very cool. I have something similar in the works. But for a different application.

nice! what's the application?

Congrats! What’s the next big feature the team has lined up?

more use-cases! custom llms are very very much under-tapped and we want to show developers more of the cool/fun stuff you can do with custom models

Can’t wait to try it

:)

I fear we all know who the star is 🐙

I really want to build a version of Zaxy for open weight models. It would be so interesting to move beyond "amnesiac with a great notebook", or at least try it. Lots more research to do.

zaxy looks cool! wonder if one could post-train a model to be effective at retrieving the right stuff from memory

That is what I'm dying to investigate!

give castform a shot then :) $50 in free credits for new users and i'm happy to get on a call to get you set up

This is awesome - congrats Girish!

thanks alex!!

sick!

thanks!! please give it a shot and let us know what you think :) new users get $50 in free credits

cool video :)

thanks khushboo!!

This is actually cool, man. I'm actually training an Omni model from scratch. It is going to be a 1 billion parameter model, and it took me three months just to get to this level right now.

ooh very cool! why from scratch?

Because at the beginning I thought of using QWEN OMNI but it was a minimum of 3 billion parameters and 1. I wanted to make it as small as possible. that's why I made it like 1 billion 2. also I trained it on regional Indian languages which qwen does not have.

ahh gotcha that makes sense!

By the way I am really looking for any feedbacks from you I have dm'ed you

great will respond!

Woohoo congrats! Love the Pokemon reference 😁

hahaha hope nintendo doesn't sue us 👀

The real unlock is agent traces as training data. Every session Fable 5 runs for you is a dataset. Castform turns your best AI outputs into a cheaper model that does the same thing.

🚀🚀🚀

Congrats @googrish !

thank daanish! really enjoyed our convo on rubrics when u came over to the office!

👏🏻 Audio models when?

soon! what would you post-train an audio model for?

Lotsa music use cases suffer from having low data volume (e.g. duet vocals, non-western music) so post training would help boost their performance using small but high quality datasets 🎵

ahh very interesting! cc @YingHang7

So excited!!!

thanks sudz!

compute is scarce and will stay scarce. smaller, specialized models is key to getting around that constraint. Good to see a platform built around that bet instead of fighting it