LoRA fine-tuning · batch inference
Get one price for your LoRA fine-tune.
Tell us the open model, the training tokens, the epochs and when you need it. A person replies by email with the day's price and how to send your dataset. A request is not an order: nothing is booked or charged.
- No account
- No card
- Not an order
- Priced on the day
How it works
Three steps. One price.
You describe the job. A person prices it on the day and replies by email. Sending a request books nothing and charges nothing.
- 01 · REQUEST
Describe the job.
Name the base model, a rough size, the epochs and when you need it. No account, no card, and no dataset yet: the form asks only for the job's shape.
- 02 · PRICED
A person prices it on the day.
Prices move with the GPU market, so every reply carries the date it was priced. Each fine-tune is compared with Together's published fine-tuning list for the same model size, tokens and epochs. If their list is lower for your job, the reply says so.
- 03 · REPLY
You decide.
The reply carries the day's price and how to send your dataset. If you accept, you pay by invoice or payment link. Nothing is charged from this page.
Fine-tuning
LoRA fine-tunes of open-weight models.
We price LoRA fine-tunes of open-weight models such as Qwen3.8 27B. Name any other model; if we cannot price it, the reply says so.
Fine-tune your base model with LoRA on your training tokens for a few epochs, needed when you need it.
- The priceOne price for the job.
It depends on the base model, the training tokens, the epochs and when you need it. The reply states one price for the job you described, and the date it was priced.
- The resultAn adapter for your base model.
A LoRA fine-tune produces an adapter for the base model you chose. Tell us in the notes how you plan to use it.
- Where it runsOn GPUs we rent.
Your job runs on GPUs we rent from cloud providers, under our own account. We do not name the providers.
- Your dataNo dataset in the form.
Do not put your dataset in the request form. The reply tells you how to send it.
Batch inference
Running batch jobs? Same request.
Batch inference on open models can be requested the same way. Name the model, a rough count of input and output tokens, and when you need it. A person replies with the day's price and how to send your input file.
FAQ
Short answers.
Is a request an order?
No. A request is not an order: nothing is booked or charged when you send it. If you accept the price in our reply, you pay by invoice or payment link. Nothing is charged from this page.
How is the price set?
A person prices each job on the day. The price depends on the base model, the training tokens, the epochs and when you need it. Prices move with the GPU market, so every reply carries the date it was priced.
For fine-tuning, we compare each job with Together's published fine-tuning list for the same model size, tokens and epochs. If their list is lower for your job, the reply says so.
Which models can I ask about?
We price LoRA fine-tunes of open-weight models such as Qwen3.8 27B. Name any other model by its Hugging Face repo id or its name. If we cannot price it, the reply says so.
What do I send now?
Only the shape of the job: the model, a rough size, the epochs and when you need it. Do not put your dataset in the form. The reply tells you how to send it.
Where does it run?
On GPUs we rent from cloud providers, under our own account. We do not name the providers, and you do not need to choose the hardware.
Is anything cheaper?
Sometimes. Free previews and promotions from other providers can beat any paid price, ours included.
Is my data certified-secure?
We do not claim SOC 2 or ISO 27001 certification, and we do not borrow a provider's certificate.
When do I hear back?
A person reads every request and replies by email from contact@computescan.ai.
Request a price
Tell us the job. A person replies.
No account and no card. A person reads every request and replies by email with the day's price and how to send your data.