The basics
What is Quantinia? ▶
A platform to finetune and run open-weight AI models on your own infrastructure. Bring your data, pick a model, and we turn it into a specialist you own - running on your hardware, not rented by the token.
What does "finetuning" mean here? ▶
Finetuning teaches a base model how your field really works. We use efficient methods - LoRA/QLoRA adapters plus SFT and RLVR - small files that specialize a model without retraining it from scratch. You can keep several adapters and swap them anytime.
Which models can I use? ▶
Any open-weight model - Qwen, Llama, Mistral, DeepSeek, Gemma and more. You pick the one that fits your task, your hardware, and your budget - compact for a single GPU, large for a cluster.
Do I need to know machine learning? ▶
No. To use a model, you just run it. To finetune one, you provide the data and the pipeline handles the rest - or, if you'd rather, we do it for you (see hosted training below).
Privacy & ownership
Where does my data go? ▶
Nowhere. The model runs on your own machine or infrastructure, so your prompts, documents, and answers never leave your network.
Do I own the result? ▶
Yes. The finetuned model and adapters are just files you own. Keep them entirely local, or store them in a private repo with access control for your org. Nothing is public unless you publish it.
Does it work offline? ▶
Yes. Once the model is on the machine, everything runs fully offline - useful for regulated work, remote sites, or anywhere data can't leave the building.
Hardware
What GPU do I need? ▶
It depends on the model size. To run a compact model, a mainstream card works. To finetune, a 24 GB card (RTX 3090/4090/5090 or Radeon RX 7900 XTX) covers many models; larger models want more memory or a cluster. NVIDIA (CUDA) and AMD (ROCm) are both first-class. No suitable card? Use hosted training.
Can I run it on CPU only? ▶
Only for the smallest models, and slowly. For anything serious a GPU is needed to run it well - that's why hosted training and hosted inference options exist for those without one.
Pricing
What does it cost? ▶
Flat and transparent - free to explore, a predictable fee for private use, and hosted training priced from your dataset. There's
never a per-token or per-seat meter. See the
Pricing page for details.
How is hosted training priced? ▶
Automatically, from your dataset. You upload it, we show an exact quote before anything runs; you approve, it trains on our GPUs and lands in your account. Bigger datasets cost more in proportion to the GPU time they need.
Why is it cheaper than a frontier subscription? ▶
Because you run it yourself. We don't rent you intelligence by the token - you own the model and it does the work on your hardware. Frontier models escalate the bill with every seat and every query; here, everyday work runs locally and you only pay a frontier model for the rare ~1% of genuinely hard cases.
For companies
Can we finetune on our internal data? ▶
Yes - that's the point. Start from an open-weight model, finetune on your handbook, codebase, tickets, or records, and keep the resulting model private. The raw data never leaves your network to do it.
We don't have the GPUs or the time. Can you do it? ▶
Yes. Beyond automated hosted training, we offer hands-on consultancy - we come in, understand your domain, build a coherent dataset, and deliver a working model. Far cheaper than hiring and retaining an ML specialist.
Getting started
Is there a library of examples to start from? ▶
A public library of published models and datasets is on the way. For now, we start your project from a strong open-weight base and build the specialized model together with you.
Do I need an account? ▶
Only to contribute or manage private work. Browsing and downloading public examples needs nothing. An account is just an email and password.