spinf

spinf, made by spinf Inc. in San Francisco, runs open-weights language models as decision models: content is read once and every question returns a probability over the answers you allow. It is offered as a hosted API (currently in beta), as the open-source spinf-serve engine under Apache 2.0, and as private deployments.

What It Is

spinf is a bulk classification and scoring service. A question is written as a template that ends where the answer goes, with a list of answers to score. The response gives a probability for each answer rather than generated text, so classification, scoring and decisions can be run over many documents. The spinf-serve engine exposes an OpenAI-compatible API and uses the same request and response format as the hosted service.

How scoring works

Each question returns three numbers: p (probability of an answer), residual (probability that went to words outside your options) and floor_p (the same answer scored on empty content, so p minus floor_p shows what the content added). Placeholders in a template expand one question into a grid, and many contents can be sent in one call. Extraction slots can decode text, numbers and amounts from the same read. Billing covers input tokens only, with no charge for the scores returned, according to the vendor.

Use cases and prompt packs

The site lists news analysis, survey and review analysis, content classification, content moderation, ticket and message routing, and prompt-injection detection. Prompt packs are ready-made, calibrated question sets, such as a moderation pack, that can run in the same call as your own questions. The vendor presents a case study in which Thinking Text built daily oil and equity news indexes on spinf.

Running it yourself

spinf-serve installs with pip or Docker and runs on one 24 GB NVIDIA GPU on Linux, with fp8 tensor cores (Ada, Hopper or Blackwell) and CUDA 13 on Python 3.12; Ampere cards use int8 weights. Published checkpoints include Gemma 4 12B and 31B, Qwen3.8-27B and a Perplexity decision model. Some models also read images, audio or video frames. The README reports its own benchmark against vLLM on one RTX 4090.