Loading…
How it works
We serve the open-weight model unmodified, so within its own context window its answers are its own — nothing of ours to explain away. Past that window we carry context by retrieval rather than by a longer prompt, and that retrieval work runs on CPU instead of on your GPU. You index your documents once, then reuse and recombine them across chats and agents with no prefill to redo. Here is the whole pipeline, in buyer terms.
You start from an open-weight model you already trust, and we serve it exactly as its authors published it: no conversion step, no retraining, no fine-tune. Nothing about the model's behaviour is ours to explain away. What changes is everything around it — the serving stack is where the work is. Today, that's one supported model.
For ML buyers — what we touch, and what we deliberately don’t.
Through the console, an API call (including from another application), or an integration, a user selects a document to be ingested, indexed, and stored — supporting multiple versions and edits. This is a one-time cost per document.
The “set it up once” promise.
The same model uses any stored document — or any combination — as part of a chat or agent: the console for testing, the API for production. Because step 2 is persisted, there is no prefill to re-run between uses, even days apart.
Where cost and latency savings land.
The model's own weights, standard tokenizers, and an OpenAI-compatible endpoint — point your existing stack at it and keep your tooling.
Index once. Reuse forever. Inside your own AWS account.
Coming soon: hosted models on lab358 Cloud, or self-hosted via AWS Marketplace.