# Shisa.AI: shisa-de-1

> Shisa AI's decision model, a fine-tune of Gemma 4 26B-A4B (about 3.8B parameters active) that reads Choice, Noul and Score answers from the option-letter distribution instead of generating text. English and Japanese, text or images; open weights plus a hosted endpoint.

- Page: https://systemonemodels.ai/shisa-ai/shisa-de-1
- API: https://api.systemonemodels.ai/v1/models/shisa-ai/shisa-de-1
- Download: `pip install systemonemodels && systemone pull shisa-ai/shisa-de-1`

## Facts

| | |
|---|---|
| Maker | Shisa.AI (https://systemonemodels.ai/shisa-ai) |
| Decides | choice, score, noul, classify |
| Architecture | shisa-de |
| Base model | google/gemma-4-26b-a4b-it |
| Parameters | 25B |
| Context | 256K tokens |
| Licence | apache-2.0 |
| Availability | Open weights + hosted API |
| Released | 2026-09-21 |
| Latest version | 2026.09 |

## Hosted API

- Provider: Shisa AI (https://platform.shisa.ai/)
- Docs: https://github.com/shisa-ai/shisa-de
- Input: — · Output: —

## Reported evaluation

Suite: JevBench v1.2 public set (231 items; 191 correct), maker's run; ECE on the hard tier. Numbers are the publisher's own.

- Decision accuracy: 82.7%
- Calibration error (ECE): 0.083
- Median latency: 20 ms
- p95 latency: 23 ms

## Model card

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

# Shisa DE-1

Shisa AI's decision model, a fine-tune of Gemma 4 26B-A4B (about 3.8B parameters active) that reads Choice, Noul and Score answers from the option-letter distribution instead of generating text. English and Japanese, text or images; open weights plus a hosted endpoint.

Shisa DE-1 takes one state and a list of named questions and returns one typed answer per question. Each answer is read from the next-token distribution at the answer position, restricted to the option letters (A upward, at most 26 per prompt); the training loss was cross-entropy over exactly that restricted distribution, on 3,632 examples (compaction, fraud detection, anti-slop and safety guardrails). The vision tower was left untouched. Serves on stock vLLM through /v1/completions with logprobs; an arbitrary chat endpoint is not enough. The maker's shisa-de Python client (Apache-2.0) wraps the readout and also talks to the hosted endpoint on the Shisa Platform (API key required; no public price found). Caveats from the maker's own card: the frozen Gemma 4 base matches or beats DE-1 on several of the maker's suites (kev-decision-v1 0.742 vs 0.755 for the base); option order changes answers (mean absolute shift 0.029, up to 11 points on an 11-option counting probe); arithmetic and counting are weak; raw probabilities are not calibrated, and the client applies a post-hoc temperature fitted by the maker (choice ECE 0.138 to 0.046 on its test split) which the maker says to refit on your own rows. Choices with more than 26 options use a chunk-and-finalist rule whose scores are never calibrated. The BF16 weights are 48 GiB; the maker also publishes an FP8-dynamic build (25.3 GiB). A DE-2 is named in the client but is not released. On third-party boards: the maker reports its own run of the JevBench v1.2 public items (below); DE-1 has no official JevBench v1.4 score, and the maker cites 37.79 on the third-party Decision Index from its own run.

## 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
- **classify** — assigns a category from a fixed taxonomy

## At a glance

| | |
|---|---|
| Parameters | 25.2B |
| Base model | `google/gemma-4-26b-a4b-it` |
| Maker | Shisa.AI |
| Released | 2026-09-21 |
| License | apache-2.0 |
| Reported accuracy | 82.7% |
| Reported latency | 20 ms p50 / 23 ms p95 per standard-tier decision, 52 / 156 ms on hard-tier items, on 2x NVIDIA H20-3e (TP2, stock vLLM), client wall time |

## Hosted API

Served by **Shisa AI**. [Get access](https://platform.shisa.ai/) · [API docs](https://github.com/shisa-ai/shisa-de).

## Get the weights

```bash
pip install systemonemodels
systemone pull shisa-ai/shisa-de-1
```

The files are served from the maker's Hugging Face repository, [`shisa-ai/shisa-de-1`](https://huggingface.co/shisa-ai/shisa-de-1), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/shisa-ai/shisa-de-1)
- [FP8-dynamic build](https://huggingface.co/shisa-ai/shisa-de-1-FP8-dynamic)
- [Python client (shisa-de)](https://github.com/shisa-ai/shisa-de)
- [JevBench run records (maker's)](https://github.com/shisa-ai/jevbench-results)
- [Shisa Platform](https://platform.shisa.ai/)

---

*This page was opened by System One for Shisa.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