# Kortexa AI: shingi-27b

> A 27B ternary decision model in one 7.2 GB GGUF: Prism ML's Bonsai 2 27B with retrained block scales. It reads the logit of every candidate answer to return a choice, a yes/no probability or an ordinal score, takes text and up to 8 images, and serves /v1/systemone locally.

- Page: https://systemonemodels.ai/kortexa-ai/shingi-27b
- API: https://api.systemonemodels.ai/v1/models/kortexa-ai/shingi-27b
- Download: `pip install systemonemodels && systemone pull kortexa-ai/shingi-27b`

## Facts

| | |
|---|---|
| Maker | Kortexa AI (https://systemonemodels.ai/kortexa-ai) |
| Decides | choice, score, noul |
| Architecture | shingi |
| Base model | prism-ml/ternary-bonsai-2-27b-gguf |
| Parameters | 27B |
| Context | 16K tokens |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-29 |
| Latest version | 3.2.0 |

## Reported evaluation

Suite: JevBench public set (231 items), the maker's own run of v3.2; also used to guide training data. Numbers are the publisher's own.

- Decision accuracy: 85.7%

- Median latency: 66 ms

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# Shingi 27B

A 27B ternary decision model in one 7.2 GB GGUF: Prism ML's Bonsai 2 27B with retrained block scales. It reads the logit of every candidate answer to return a choice, a yes/no probability or an ordinal score, takes text and up to 8 images, and serves /v1/systemone locally.

Kortexa trained only the per-block scales of Bonsai 2 27B (ternary, built from Qwen3.8-27B); the ternary weights are unchanged and there is no adapter. It needs the Prism fork of llama.cpp and an NVIDIA GPU with 20 GB or more, or Apple Silicon with Metal (about 1.2 s per short decision on an M4 Pro). The local server answers /v1/systemone and SGLang's /v1/decisions shape, without authentication. Image input uses Prism ML's vision projector unchanged, and image scores are zero-shot. On the maker's own runs of the v3.2 weights: JevBench public 85.7% (231 items, a third-party benchmark) and hard 72.1%, DecisionBench 75.3%, This/That 67.9%; the maker notes these suites also guided the choice of training data, so they are development results. No calibration error is reported; the shipped calibration kept temperature 1.0. English only. About a third of the training tokens are synthetic, including Mapika's Apache-2.0 decider teacher data.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 27B |
| Base model | `prism-ml/Ternary-Bonsai-2-27B-gguf` |
| Maker | Kortexa AI |
| Released | 2026-09-29 |
| License | apache-2.0 |
| Reported accuracy | 85.7% |
| Reported latency | 66 ms median per short yes/no decision on an RTX PRO 6000, 85 ms on an RTX 4090, one request at a time |

## Get the weights

```bash
pip install systemonemodels
systemone pull kortexa-ai/shingi-27b
```

The files are served from the maker's Hugging Face repository, [`kortexa-ai/shingi-27b`](https://huggingface.co/kortexa-ai/shingi-27b), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/kortexa-ai/shingi-27b)
- [Code and server](https://github.com/kortexa-ai/shingi-27b)
- [Base model (Bonsai 2 27B)](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf)

---

*This page was opened by System One for Kortexa AI, who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models