nebula@atheron:/batch$
// Batch inference
Batches, done for you.
Embeddings, labels, transcripts, captions.
Upload your items, pick a model from our catalogue, and get one result file back. We split the work across GPUs on the Constellation network, check every part, and put it back together in order.
// What you can run
Four jobs to start
Each runs on open models we have checked, signed and pinned, so the same input gives you the same kind of answer every time.
Embeddings
Turn text into vectors for search, clustering and recommendations. Small, base and large models, normalised and ready for your vector store.
Classification
Sort text into your labels: topics, sentiment, intent or spam. You get a label and a score for every item.
Transcription
Turn recorded audio into text with timings, in the languages the model supports.
Captions
Describe images in a sentence, for alt text, search and cataloguing.
// Checked work
Every part is verified
A batch is only as good as its worst part, so no part is trusted on one machine's word.
Run twice
Each shard on member devices runs on at least 2 machines owned by different people. Their answers have to match.
Settled by us
When answers differ, an Atheron house node runs the shard itself and decides. A machine that was wrong is struck and stops getting work.
Spot checks
A share of agreed answers is run again on our own hardware at random, so nobody can learn when they are being checked.
Matched to the hardware
A shard goes only to a GPU that has the memory and the drivers the model needs. Nothing is guessed at run time.
// How a batch runs
From upload to result file
You make one call. Everything after it happens on our side.
- 1
Submit
Send your items and the model you want with one API call, or upload a file in the developer portal.
- 2
Split
We cut the batch into shards sized for the GPUs that can run that model.
- 3
Run and check
Shards run on Constellation GPUs, each one verified by a second machine or by an Atheron house node.
- 4
Collect
We put the answers back in your order, store one result file for you, and call your webhook when it is ready.
Questions
- Is my data sent to strangers' computers?
- Only if you allow it. A batch is private by default and runs only on Atheron house nodes or AWS GPUs. Mark a batch shareable to let it use member devices, which is cheaper. Never mark personal or confidential data shareable.
- How long does a batch take?
- It depends on its size and the model. You can watch each batch's progress in the portal or through the API, and a webhook tells you when it is done.
- Which models can I use?
- The models in our catalogue, listed in the docs with their licence and what they are good at. We add models after checking them.
- How long are results kept?
- 30 days by default, then deleted. You can choose a shorter window, up to 90 days at most.
Run your first batch
When Nebula opens, you create your team in the developer portal, pass a quick business check, add a card, and send a batch with your key.