---
license: apache-2.0
base_model:
- jaredpalmer/kev-0.8b
- Qwen/Qwen3.5-0.8B-Base
- jaredpalmer/kev-4b
- Qwen/Qwen3.5-4B-Base
- jaredpalmer/kev-9b
- Qwen/Qwen3.5-9B-Base
library_name: onnx
tags:
- ollaya
- onnx
- decision-model
- system-one
pipeline_tag: text-classification
---

# kev for Ollaya

[Ollaya](https://github.com/ollaya-dev/ollaya) package of **[jaredpalmer/kev-0.8b](https://huggingface.co/jaredpalmer/kev-0.8b)** and **[Qwen/Qwen3.5-0.8B-Base](https://huggingface.co/Qwen/Qwen3.5-0.8B-Base)** and **[jaredpalmer/kev-4b](https://huggingface.co/jaredpalmer/kev-4b)** and **[Qwen/Qwen3.5-4B-Base](https://huggingface.co/Qwen/Qwen3.5-4B-Base)** and **[jaredpalmer/kev-9b](https://huggingface.co/jaredpalmer/kev-9b)** and **[Qwen/Qwen3.5-9B-Base](https://huggingface.co/Qwen/Qwen3.5-9B-Base)** by Jared Palmer (adapter and pointer head) and the Qwen team (base model).
Ollaya runs open decision models locally, the way Ollama runs LLMs: typed questions in,
calibrated answers out, behind a TypeSafe-compatible API.

```sh
ollaya run kev
```

## What is in this repository

This repository holds only the files Ollaya derives, with no weights. Each graph is an ONNX export of the
original model whose weights **reference the authors' own weight files by byte offset**,
so `ollaya pull` downloads the weights from the upstream repositories, unmodified and pinned to a
commit, and verifies their sha256.

| Tag | Upstream | Files |
|---|---|---|
| `kev:0.8b` | [jaredpalmer/kev-0.8b@9a45d25](https://huggingface.co/jaredpalmer/kev-0.8b/tree/9a45d25eb2ab761841196625383fa1dff0e56c1e), [Qwen/Qwen3.5-0.8B-Base@dc7cdfe](https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/tree/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68) | `0.8b/model-fp32.onnx`, `0.8b/decision.json`, `0.8b/calibration.json` |
| `kev:4b` | [jaredpalmer/kev-4b@139fdd9](https://huggingface.co/jaredpalmer/kev-4b/tree/139fdd94f1b6a6ad80cc15e08fcb99cac885a101), [Qwen/Qwen3.5-4B-Base@1001bb4](https://huggingface.co/Qwen/Qwen3.5-4B-Base/tree/1001bb4d826a52d1f399e183466143f4da7b741b) | `4b/model-fp32.onnx`, `4b/decision.json`, `4b/calibration.json` |
| `kev:9b` | [jaredpalmer/kev-9b@2629c06](https://huggingface.co/jaredpalmer/kev-9b/tree/2629c06a5aeb0feb3b9783bafed17ed8f39ecf5c), [Qwen/Qwen3.5-9B-Base@68c46c4](https://huggingface.co/Qwen/Qwen3.5-9B-Base/tree/68c46c4b3498877f3ef123c856ecfde50c39f404) | `9b/model-fp32.onnx`, `9b/decision.json`, `9b/calibration.json` |

Each tag has an fp32 graph, used on CPU and GPU. Each tag also has `decision.json` (sequence layout, special tokens) and
`calibration.json` (temperatures).

## Parity

Ollaya's Rust runtime matches upstream Kev (PyTorch fp32) exactly on 480 questions from 117 requests per checkpoint, and rejects the same 16 requests upstream rejects. The token rows and option positions are identical, and so is the decision on every question. Probabilities are within 2.8e-6 (0.8b), 3.1e-5 (4b) and 3.8e-6 (9b), on CPU and CUDA, and the TypeSafe answers equal upstream's to its 4-decimal rounding.

## License

Same as the upstream model (Apache-2.0). Ollaya itself is Apache-2.0.