FAQ

I need help or found a bug. Who can I ask?

Write to us at support@spinf.com. Our team is in California, but we try our best to get back to you as soon as possible, whatever the time. Include the request_id of the call if you have one: it is in every response and error, and in the X-Request-Id header.

Nothing works!

spinf is in beta, so it can happen at times: we may be briefly down or upgrading the system. Drop us an email anyway. Any input, critique or reminder is appreciated.

Why do I write a template with answers instead of asking a free-form question?

We tested several approaches, and a question written as a template that ends where the answer goes, with the answers you choose (" rise", " fall"), proved much more controllable and predictable than free-form questions or generated text. You decide exactly which words are scored and where the model reads them, which gives the level of control that fine-grained analysis needs. See how a question works.

Are the scores deterministic?

Yes. The engine is batch-invariant: a content gets the same scores whether it is scored alone or packed with thousands of others, bit for bit, and repeating a call returns the same numbers. Scores do change when the engine version, the weights or the GPU architecture change; the system_fingerprint of each response identifies that configuration, so you can tell whether two responses are directly comparable, and recalibrate when it changes (see Calibration). The open-source engine gives the same scores as the hosted service on GPUs with fp8 tensor cores (Open source).

Are images, audio or video supported?

Yes, but they are not generally available: images, video (sent as frames) and audio are in preview and enabled per account on request (Images, video and audio). If you want to work with us to test a use case, let us know: we may open these up to selected partners over time.

Can spinf make decisions in real time?

Yes, on a reserved endpoint. The on-demand service is built for throughput (batches, backfills, archives), and the first request after an idle period may take a few seconds. For live decisions we reserve capacity for you, always warm and sized to your traffic (Limits & pricing).

Can I run spinf myself?

Yes. The engine behind this API, spinf-serve, is open source under the Apache 2.0 license: pip install "spinf-serve[server]", or the container image ghcr.io/spinfinc/spinf-serve, serves the same models on one 24 GB NVIDIA GPU, with the same request and response, so code, thresholds and calibrations move between your server and the hosted API by changing the base URL and the key. See Open source and the repository at github.com/spinfinc/spinf-serve.

Are private deployments supported?

Yes, although they are not generally available yet. If you have a large classification project, let us know: we can provision dedicated spinf capacity that meets your processing timeline.

Which models do you run?

Google DeepMind's open-weights Gemma 4 models, optimized by us for scoring at scale: spinf-12b is Gemma 4 12B, and spinf-31b (on request) is Gemma 4 31B. The API accepts either name: "model": "google/gemma-4-12B" works as well as "model": "spinf-12b". See Models. The checkpoints we serve are public on the Hugging Face Hub under spinf/, for the open-source engine. Gemma is a trademark of Google LLC; the Gemma 4 weights are released under the Apache 2.0 license.

Do you publish benchmarks?

Yes. spinf-benchmarks on GitHub has the test sets (synthetic and public), the exact questions and example sets, the results per model, and the code to re-run everything with your own API key. The figures by use case are on the use-case pages.

Can you support a specific open-weights or custom model?

Possibly. Supporting a model means adapting it to our inference engine, calibrating it, and adapting it to several kinds of hardware, and in our testing not every model is suitable or can be accelerated to the speed we need. If you have a suggestion or a request, write to us.

I need professional support to build my own integration. Can you help?

Yes. We can most likely help you directly, or point you to one of our integration partners.