YOUR OWN AI MODEL FOUNDRY
From a domain brief to a deployed model.
Gro turns your corpus into a real, owned model — pretrained, fine-tuned, aligned, evaluated and served as a managed API or in your own AWS account. Every step is proposed for your approval, grounded in your data, and kept inside EU residency.
Build your first model in days, one approved step at a time.
5
training methods
Yours
weights, owned outright
EU
data residency
Full
provenance on every run
HOW IT WORKS
Four approved steps to a real model.
01
Brief the domain
Describe the problem and point Gro at your corpus. It scopes the dataset, tokenizer and architecture.
02
Approve the plan
Gro proposes a training method, budget and estimate. Nothing runs until you approve it.
03
Train on real GPUs
Jobs run on managed SageMaker with a live loss curve, per-step metrics and full provenance.
04
Evaluate & deploy
Score against your eval suite, then deploy a serverless endpoint or download the weights.
FIVE REAL WAYS TO BUILD
Not a prompt wrapper. A real foundry.
Gro supports the full modern model-building lifecycle — each method is a genuinely distinct objective, not a relabelled call.
01
From scratch
Pretrain a new decoder on your corpus when you need a model that is entirely your own.
02
Continued pretraining
Adapt an existing checkpoint to your domain with more of your unlabelled data.
03
Supervised fine-tuning
Teach instruction-following on {prompt, response} pairs with prompt-masked loss.
04
Preference optimization
Align behaviour with DPO on {chosen, rejected} triples against a frozen reference.
05
Distillation
Compress a large teacher into a smaller, cheaper student that keeps the capability.
06
Not sure which?
Gro proposes the right method for your goal and data — you approve before a single GPU spins up.
GOVERNED BY DESIGN
Built for the boardroom and the AI Act.
Export the weights anytime
Download the trained model and run it anywhere — no lock-in, no per-seat rent on your own IP, no ongoing charge from us.
EU data residency
Corpus, training and inference stay in-region. Designed for the AI Act from day one.
Propose, then approve
Every cost-bearing action is proposed first with its real credit cost. A human approves before a single GPU spins up.
Full provenance
Every run records its method, data, metrics and resulting weights — auditable end to end.
DEPLOY & OWN IT YOUR WAY
Three ways to run your model — you choose.
The weights are yours the moment training finishes. Run them on our managed infrastructure, take them and self-host, or point Gro at your own AWS account.
Managed & metered
Deploy a serverless endpoint on our infrastructure in one click. You're billed cost-plus in credits — the real AWS cost of each call plus a small platform margin, metered per request so you only pay for what you run. Idle endpoints auto-retire so a forgotten deployment never keeps billing.
Export & self-host
Download the trained weights with a generated PyTorch and SageMaker run recipe. Run them on your own laptop, server or cloud with zero ongoing charge from us — you own the model outright.
Bring your own AWS
Connect a cross-account IAM role and Gro runs training, deployment and inference inside your own AWS account, on your own bill. We never store your keys, and we waive cost-recovery credits entirely.
CREDITS, NOT SEATS
Cost-plus credits. Pay for the compute you approve.
Start free with 20 credits a month. Upgrade to Professional or Scale for a bigger monthly allowance. Every charge recovers the real AWS cost plus a small platform margin — training is refunded if a job fails, and live endpoints are metered per call.
Free
Explore the studio, build datasets and run a first small job.
20 credits / month
+ 150 welcome credits to complete your first full run
Professional
For iterating across projects — fine-tuning, evaluation and deployment.
125 credits / month
Scale
Production workloads with room for distillation, DPO and served endpoints.
1,000 credits / month
The Free plan includes 20 credits every month — no card required to explore. Bring your own AWS and cost-recovery credits are waived entirely.
FAQ
Questions, answered.
- Do I really own the model?
- Yes. The trained weights are yours to download, deploy or delete at any time — with a generated run recipe so they work on your own hardware or cloud. Gro orchestrates; you own the output.
- What does a credit pay for?
- Cost-bearing steps — synthetic data generation, dataset builds, training runs, evaluations, deployments and live inference calls. Pricing is cost-plus: each charge recovers the real AWS/compute cost plus a small platform margin. Briefs, configuration and refreshes are free.
- How is inference billed?
- Managed endpoints are metered per call. Each request is charged in credits based on its actual serverless compute (memory × duration) plus margin, so you only pay for what you run. Idle endpoints auto-retire to stop surprise cost.
- Can I run models in my own AWS account?
- Yes. Connect a cross-account IAM role and Gro runs training, deployment and inference inside your own AWS account on your own bill. We never store your keys, and cost-recovery credits are waived — you pay AWS directly.
- Where does my data live?
- Inside EU residency. Your corpus is never used to train anyone else's model. Bring your own AWS to keep everything in your own account and region.
- Which methods are supported?
- From-scratch pretraining, continued pretraining, supervised fine-tuning, preference optimization (DPO) and distillation — each a real, distinct objective.
- Are credits refunded on failure?
- Yes. Training and other launched jobs are refunded if the run fails or is cancelled — a failed launch costs nothing.
AI THAT REACHES PRODUCTION
Weeks to a working model, not months to a slide deck.
Bring a domain brief. Leave with a deployed, owned model.