Macaw/README.md

---
language:
  - en
license: other
library_name: transformers
license_name: lfm-open-license-v1.0
license_link: https://huggingface.co/LiquidAI/LFM2.5-2.6B/raw/main/LICENSE
tags:
  - lfm2
  - macos
  - agent
  - tool-calling
  - on-device
pipeline_tag: text-generation
base_model: LiquidAI/LFM2.5-2.6B
---

# 🦜 Macaw

**The assistant your Mac should have shipped with.**

Ask for something in plain English and Macaw does it — sends the email, finds the
file, reads the PDF, tells you what's eating your battery. It runs entirely on
your Mac. Nothing is uploaded, nothing is logged, and there is no account to
make.

Available today. Not a waitlist.

👉 **[badtheorylabs.com/macaw](https://badtheorylabs.com/macaw)**

---

## What you can say to it

```
"email ada the q3 numbers and tell her i approved"
"why is my mac slow?"
"read ~/Documents/contract.pdf and summarise it in two sentences"
"take a screenshot, make a folder called Shots, and move it there"
"what's on my screen?"
"am i free tomorrow afternoon?"
"find my biggest files"
```

It chains steps on its own. *"Take a screenshot, make a folder, move it there"*
is three actions from one sentence, and it checks each one worked before moving
to the next.

---

## What it can reach

**97 tools**, each one hand-written and tested — not a scrape of every command
macOS exposes.

| | |
|---|---|
| ✉️ **Mail** | send, search, reply, unread counts, senders |
| 📅 **Calendar** | create events, today's schedule, free slots, next meeting's video link |
| 📁 **Files** | Spotlight search, open, move, folder sizes, biggest files, old downloads |
| 📝 **Notes & Reminders** | create, append, search, complete, overdue |
| 🎵 **Music** | play, pause, skip, by artist, by playlist, shuffle |
| 🌐 **Safari & Chrome** | current tab, all tabs, open, search |
| 🖥️ **System** | battery, disk, wifi, brightness, dark mode, volume, focus, lock, sleep |
| 👀 **Screen** | read what's on screen, record it, screenshot it |
| 📄 **Documents** | read and summarise PDFs, Word files, code, plain text |
| ⚡️ **Diagnostics** | what's using CPU, what's eating memory, battery health |

---

## Private by construction

Not a policy — an architecture. The model weighs 1.5 GB and lives on your
machine. There is no server to send anything to.

- No cloud, no API keys, no account
- No telemetry, no analytics, no crash reporting
- Works on a plane
- Your mail, files and calendar never leave the device

---

## Specs

| | |
|---|---|
| **Size** | 2.7B parameters · 1.5 GB on disk (4-bit) |
| **Memory** | ~2 GB while running |
| **Speed** | ~1.2 s per request on an M2 |
| **Context** | 128K tokens |
| **Requires** | Apple Silicon (M1 or later), macOS 14+ |

**Two builds:**

- **[`Macaw-4bit-MLX`](https://huggingface.co/badtheorylabs/Macaw-4bit-MLX)** —
  1.5 GB, what the app runs. Start here.
- **`Macaw`** (this repo) — 5.4 GB BF16, for fine-tuning or serving on a GPU.

---

## Run it yourself

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("badtheorylabs/Macaw")
model = AutoModelForCausalLM.from_pretrained("badtheorylabs/Macaw",
                                             dtype=torch.bfloat16)

messages = [
    {"role": "system", "content": "You are Macaw, an on-device assistant running on this Mac."},
    {"role": "user", "content": "what's my battery at?"},
]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True,
                                 return_dict=True, return_tensors="pt")
out = model.generate(inputs["input_ids"], max_new_tokens=128)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

Pass your tools with `tools=[...]` and it replies with a call:

```
<|tool_call_start|>[battery_status()]<|tool_call_end|>
```

The app in [github.com/Badtheorylabs/Macaw](https://github.com/Badtheorylabs/Macaw)
handles the rest — parsing, running, and asking before anything destructive.

---

## Measured

On an Apple M2, 4-bit build:

| | |
|---|---|
| Tool-call accuracy | **10 / 10** on the capability suite |
| Request latency | **1.21 s** mean · 0.77 s best |
| Decode | ~40 tok/s |

Method and harness are in the [GitHub repo](https://github.com/Badtheorylabs/Macaw).

`model.safetensors` SHA-256
`966d206651b96446f6a4c415d118049f478a8df79be3fa45bfd04959a4d68be0`

---

## License

Weights derive from [LFM2.5-2.6B](https://huggingface.co/LiquidAI/LFM2.5-2.6B)
under the [LFM Open License v1.0](https://huggingface.co/LiquidAI/LFM2.5-2.6B/raw/main/LICENSE).
Redistribution keeps that license and Liquid AI's attribution. Commercial use is
free below $10M annual revenue. The app and tooling are MIT.

© 2026 Bad Theory Labs