| --- |
| title: README |
| emoji: 🎯 |
| colorFrom: gray |
| colorTo: blue |
| sdk: static |
| pinned: false |
| --- |
| |
| # spinf |
|
|
| **Decision models: one input read once, hundreds of questions answered as probabilities.** |
|
|
| spinf turns open models into decision models. Instead of generating text, the model scores: a piece of content is read |
| once and every question about it is answered in the same pass with a calibrated, batch-invariant probability. The |
| checkpoints here are the ones [spinf-serve](https://github.com/spinfinc/spinf-serve), our open-source inference engine, |
| serves: quantized and calibrated by us, with the quantization parameters travelling with the weights, so the same model |
| gives the same score on your GPU and on the hosted API at [spinf.com](https://spinf.com). |
|
|
| | model | based on | one card of | reads | |
| |---|---|---|---| |
| | [spinf/gemma-4-12b-fp8](https://huggingface.co/spinf/gemma-4-12b-fp8) (default) | Gemma 4 12B, FP8 with static activation scales | 24 GB (RTX 4090 / 5090, L4, L40S, RTX 6000 Ada, H100) | text, images, audio, video frames | |
| | [spinf/gemma-4-31b-fp8](https://huggingface.co/spinf/gemma-4-31b-fp8) | Gemma 4 31B, FP8 static | 48 GB (RTX 6000 Ada, L40S, RTX PRO 6000) | text, images, video frames | |
| | [spinf/qwen3.8-27b-int4](https://huggingface.co/spinf/qwen3.8-27b-int4) | Qwen3.8-27B, GPTQ int4 | 24 GB | text, images, video frames | |
| | [spinf/pplx-decider-v1-27b-int4](https://huggingface.co/spinf/pplx-decider-v1-27b-int4) | pplx-decider-v1-27b, GPTQ int4: a decision model with a readout head | 24 GB | text, with its own packs | |
|
|
| ```sh |
| pip install "spinf-serve[server]" |
| spinf-serve serve spinf/gemma-4-12b-fp8 # downloads the checkpoint on first start; an OpenAI-compatible server on port 8000 |
| # or the image: docker run --gpus all -p 8000:8000 -v spinf-cache:/root/.cache ghcr.io/spinfinc/spinf-serve |
| ``` |
|
|
| - **[spinf-serve](https://github.com/spinfinc/spinf-serve)**: the engine and server (Apache 2.0), the quantization |
| recipes that produced these checkpoints, the parity fixture, the docs. |
| - **[spinf-prompts](https://github.com/spinfinc/spinf-prompts)**: ready-made, calibrated question sets (content |
| moderation, triage, products, reviews, brand, entertainment) the server preloads; community packs load from any |
| repository, such as [thinkingtext's news packs](https://github.com/thinkingtext/spinf-news-pack). |
| - **[spinf-client](https://github.com/spinfinc/spinf-client)**: the Python client, for your own server or the hosted API. |
| - **[Agent skills](https://github.com/spinfinc/skills)** for Claude Code and other coding agents. |
|
|
| Each model card carries its recipe and licence: Gemma 4 under Google's terms, Qwen3.8 and pplx-decider under theirs, |
| the quantization and calibration under Apache 2.0. |
|
|
