How to use from the
Use from the
MLX library
# Make sure mlx-lm is installed
# pip install --upgrade mlx-lm

# Generate text with mlx-lm
from mlx_lm import load, generate

model, tokenizer = load("OsaurusAI/Osaurus-AppleScript-16B-A4B-JANG_4M")

prompt = "Write a story about Einstein"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True
)

text = generate(model, tokenizer, prompt=prompt, verbose=True)

Osaurus

AppleScript-16B-A4B-JANG_4M

The flagship tool-calling model for macOS AppleScript & agentic computer-use. Given a run_applescript tool, it emits a structured tool call with correct AppleScript to drive macOS โ€” app automation (Safari, Finder, Mail, Notes, Calendar, Reminders, System Events), system control, clipboard, screenshots, ASObjC, and do shell script โ€” ready to execute in an agent loop. Without a tools spec, it writes AppleScript directly (hybrid).

Built by Osaurus. Derived from gemma-4-26B-A4B (MoE): expert-pruned to a hyper-focused 16.1 B (~4 B active), fine-tuned for AppleScript + tool-calling, quantized to JANG_4M (8-bit attention, 4-bit routed experts) for fast on-device inference via MLX.

Parameters 16.1 B total ยท ~4 B active (MoE)
Quant JANG_4M โ€” 8-bit attention, 4-bit routed experts
Size ~11 GB
Tool-calling native (gemma tool-call format), run_applescript
Runtime MLX (Apple Silicon) / Osaurus

Benchmark โ€” base vs final (held-out executable AppleScript bench, 87 tasks)

Tool-call emission = emits a valid run_applescript call. Compile = valid AppleScript. Exec = it runs and returns the correct result.

Tool-call emission Compile Exec
Base gemma-4-26B-A4B โœ— (writes raw, no tool calls) ~88% unreliable
AppleScript-16B-A4B-JANG_4M 100% โญ 100% โญ 84% โญ

Highest-quality AppleScript model in the Osaurus line: 100% structured tool-call emission, 100% valid AppleScript, 84% executable correctness.

Usage (MLX, tool-calling)

from mlx_lm import load, generate
model, tok = load("OsaurusAI/AppleScript-16B-A4B-JANG_4M")
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":"Create a note titled 'Groceries' in Notes."}]
prompt = tok.apply_chat_template(msgs, tools=tools, add_generation_prompt=True, tokenize=False)
print(generate(model, tok, prompt=prompt, max_tokens=300))

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

  • AppleScript-16B-A4B-JANG_4M (this) โ€” flagship quality (~11 GB).
  • AppleScript-8B-JANG_4M โ€” fast/small tier (~5.6 GB).

License

Inherits the Gemma license โ€” review and comply with the base model's terms.


Made by Osaurus ยท contact eric@osaurus.ai

Downloads last month
388
Safetensors
Model size
16B params
Tensor type
F16
ยท
U32
ยท
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for OsaurusAI/Osaurus-AppleScript-16B-A4B-JANG_4M

Finetuned
(253)
this model