README.md

---
license: other
license_name: zyphra-zaya
language:
- en
library_name: mlx
tags:
- applescript
- macos
- computer-use
- agent
- tool-calling
- function-calling
- mlx
- moe
- osaurus
base_model: Zyphra/ZAYA1-8B
pipeline_tag: text-generation
---

![Osaurus](./osaurus-x-banner.png)

# Osaurus-AppleScript-8B-JANG_6M

**A fast, small tool-calling model for macOS AppleScript & agentic computer-use.** Given a `run_applescript`
tool, it emits a structured **tool call** with valid AppleScript to drive macOS — app automation
(Safari, Finder, Mail, Notes, Calendar, System Events), system control, clipboard, screenshots, and
`do shell script` — ready to execute in an agent loop. Without a tools spec, it writes AppleScript
directly (hybrid).

Built by **[Osaurus](https://osaurus.ai)**. Base: **Zyphra/ZAYA1-8B** (MoE). Quantized with **JANG**
(6-bit affine, MLX-native) for fast on-device inference via
[MLX](https://github.com/ml-explore/mlx). Verified in the Osaurus runtime.

| | |
|---|---|
| Parameters | 8.84 B (MoE, 16 experts) |
| Quant | **JANG 6-bit** — 6-bit affine weights + 8-bit embeddings (MLX `mx.quantize`, group 32) |
| Size | ~7.4 GB |
| Tool-calling | native (`<zyphra_tool_call>`), `run_applescript` |
| Runtime | MLX (Apple Silicon) / Osaurus |

## Benchmark — base vs final (held-out executable AppleScript bench, 87 tasks)

*Tool-call emission* = emits a `run_applescript` call. *Compile* = the emitted script is valid
AppleScript. *Exec* = it runs and returns the correct result (scored on the pure-computational subset
with a deterministic answer).

| | Tool-call emission | Compile | Exec |
|---|---|---|---|
| Base (Zyphra/ZAYA1-8B) | ✗ (writes raw text, no tool calls) | 28.9% | 30.0% |
| **Osaurus-AppleScript-8B-JANG_6M** | **100%** ⭐ | **80%** ⭐ | **75%** ⭐ |

The fine-tune teaches structured tool-calling **and** valid AppleScript (base does neither). Typical
app-automation tool-calls sit at the compile tier; the exec column is the hardest computational subset
(string/number algorithms, shell, coercions).

## Usage (MLX, tool-calling)

```python
from mlx_lm import load, generate
model, tok = load("OsaurusAI/Osaurus-AppleScript-8B-JANG_6M")
tools = [{"type":"function","function":{"name":"run_applescript",
  "description":"Execute AppleScript on macOS and return its output.",
  "parameters":{"type":"object","properties":{"script":{"type":"string"}},"required":["script"]}}}]
msgs = [{"role":"user","content":"Get the URL of the front Safari tab."}]
prompt = tok.apply_chat_template(msgs, tools=tools, add_generation_prompt=True, tokenize=False)
print(generate(model, tok, prompt=prompt, max_tokens=300))
# -> <zyphra_tool_call><function=run_applescript><parameter=script> ... </parameter>...
```

Your agent parses the `run_applescript` tool call, runs the script (e.g. via `osascript`), and feeds
the result back. (Omit `tools=` to get raw AppleScript instead.)

## Tiers

For higher accuracy use **`OsaurusAI/Osaurus-AppleScript-16B-A4B-JANG_4M`**. This 8B is the fast/small
tier for low-latency on-device automation.

## License

Inherits the **Zyphra/ZAYA1-8B** license — review and comply with the base model's terms.

---
Made by Osaurus · contact **eric@osaurus.ai**