Superclouds
Create account

SuperFission AI

Deploy a model for inference, fine-tune a model on your data, or train one from scratch.

SuperFission AI runs AI models for you. You can deploy a Hugging Face model as an inference service, fine-tune a model on your own dataset, or run a full training job. You'll find it in SuperFission AIInference & Training.

Before you start

Your account must be set up first: profile, payment method, and at least one prefix that is attached to you. Until then, the screen shows Please complete setting up your account. with a short checklist. See Set up your account.

Choose a prefix

  1. Open SuperFission AIInference & Training.
  2. Pick a prefix in the Select Prefix ... list.

Your services and jobs appear, with + New and Refresh at the top.

Your services and jobs

The table shows each item's Name, Type, Status, Model and Region. Use the buttons above it to show All, or only 🤖 Inference, 🔧 Fine-tune or 🏋️ Training.

StatusMeaning
CreatingBeing set up, or the job is running.
ActiveThe service is up, or the job has finished.
FailedIt didn't succeed. Check the settings and create it again.

The list doesn't refresh on its own. Click Refresh to see the latest status.

When there's nothing yet, the table says No services yet and invites you to click + New. If you see Failed to load services, click Refresh; if it keeps happening, sign in again.

You can't stop, cancel or delete a service or job from the console. Contact support if you need one removed.

Chat, Logs and Download are coming soon

The Actions column can show 💬 Chat, 📋 Logs and ⬇ Download. They aren't available yet. Until then, contact support if you need a job's logs or its output model.

Deploy a model for inference

An inference service runs a model so it can answer prompts.

  1. Click + New.
  2. Click the Inference card. The Deploy Inference Service form appears below.
  3. Fill in the form (details below).
  4. Click 🚀 Deploy.

You're taken back to the list, where the new service shows Creating. Click Refresh until it is Active.

FieldWhat to enter
Service NameA name you'll recognise, e.g. my-llama-service. Required.
RegionWhere the service runs. Required.
HuggingFace Model NameThe model's Hugging Face id, e.g. mistralai/Mistral-7B-Instruct-v0.2. Required.
Max New TokensThe longest answer the model may give. Starts at 512; use 64 to 4096.
TemperatureHow varied the answers are. Starts at 0.7; use 0.01 to 2.0. Lower is more predictable.
DescriptionOptional.

The limits aren't enforced by the form, so stay inside them yourself.

If you see "Please fill in Service Name, Region and Model Name.", complete those fields. "Failed to deploy" or another message means the service wasn't created: check that the model name is correct and try again. "Request failed. Please try again." means the console couldn't reach the service: check your connection and try again.

Fine-tune a model

Fine-tuning adapts an existing model to your own data.

  1. Click + New.
  2. Click the Fine-tune card. The Fine-tune a Model form appears.
  3. Fill in Job Name, Region and Base Model (HuggingFace), for example meta-llama/Llama-3.2-1B.
  4. Pick a Training Method (details below).
  5. In Dataset, choose your file. It uploads straight away; wait for the line under it to show a tick with the file name and size.
  6. Type an Output Model Name for the result.
  7. Adjust Epochs, Learning Rate and Batch Size if you need to.
  8. Click 🔧 Start Fine-tune.

You're taken back to the list, where the job shows Creating. Click Refresh to follow it.

FieldWhat to enter
Training MethodLoRA (recommended, the default), QLoRA (4-bit, uses less GPU memory) or Full Fine-tune (trains every weight, needs much more GPU).
DatasetA .json, .jsonl, .csv or .txt file, up to 500 MB.
EpochsHow many times to go through the dataset. Starts at 3; use 1 to 50.
Learning RateStarts at 0.0002.
Batch SizeStarts at 2; use 1 to 32.
DescriptionOptional.

Train a model from scratch

Full training trains every weight of the model. It needs a lot of GPU memory and takes longer than fine-tuning.

  1. Click + New.
  2. Click the Full Training card. The Full Model Training form appears.
  3. Fill in Job Name, Region and Base Model (HuggingFace), for example gpt2.
  4. In Dataset, choose your file (.json, .jsonl, .csv or .txt, up to 500 MB) and wait for the tick.
  5. Type an Output Model Name.
  6. Adjust Epochs (starts at 5, use 1 to 100), Learning Rate (starts at 0.0001) and Batch Size (starts at 4, use 1 to 32) if you need to.
  7. Click 🏋️ Start Training.
Don't leave the numbers empty

If you clear Epochs, Learning Rate or Batch Size, the job uses 3, 0.0002 and 2 instead of the training defaults. Type a value in each.

If a job doesn't start

  • "Please upload a dataset first." Choose a dataset file and wait for the tick before you start.
  • "✗ Upload failed" or "✗ Upload error". The dataset didn't upload. Check that it is one of the accepted types and no larger than 500 MB, then choose it again.
  • "Please fill in all required fields." Fill in the job name, region, base model and output model name.
  • "Failed to start job" or another message. The job wasn't created. Check the base model name and try again.

To go back to the list without creating anything, click Cancel on the form.