Metadata-Version: 2.4
Name: alakazam-mira-mini
Version: 0.1.1
Summary: Play MIRA Mini — a neural world model of car soccer — locally on your own GPU (CUDA, Apple Silicon).
Project-URL: Homepage, https://alakazam.gg/mira
Project-URL: Technical report, https://alakazam.gg/mira
Project-URL: Weights (Hugging Face), https://huggingface.co/alakazamworld
Project-URL: Upstream, https://github.com/mira-wm/mira
License: MIT
Requires-Python: >=3.10
Requires-Dist: einops>=0.8
Requires-Dist: fastapi>=0.110
Requires-Dist: huggingface-hub>=0.23
Requires-Dist: numpy
Requires-Dist: omegaconf>=2.3
Requires-Dist: opencv-python-headless>=4.9
Requires-Dist: pillow
Requires-Dist: pydantic>=2.7
Requires-Dist: termcolor
Requires-Dist: torch>=2.6
Requires-Dist: torchmetrics
Requires-Dist: tqdm
Requires-Dist: uvicorn[standard]>=0.27
Requires-Dist: websockets>=12
Description-Content-Type: text/markdown

# alakazam-mira-mini

Play **MIRA Mini** — a neural world model of car soccer, every frame generated live by the
model — locally on your own GPU. No cloud, no account; after the first weight download
everything runs on your machine.

```
pip install alakazam-mira-mini
mira-mini play
```

MIRA Mini is our from-scratch reproduction of the MIRA recipe
([General Intuition](https://www.generalintuition.com/) × [Kyutai](https://kyutai.org/),
with Epic Games), compressed — fewer diffusion steps, a smaller student model, a compact
decoder — until it runs on consumer hardware. The full story, with measurements, is in the
[technical report](https://alakazam.gg/mira).

## What you need

- An NVIDIA GPU (CUDA), or a Mac with Apple silicon (M1 or newer). Plain CPU is refused —
  too slow to be fun.
- Disk for the weights, downloaded once from Hugging Face: ~5 GB for the 364M model,
  ~12 GB for the 1B.
- The weight repositories unlock at launch; until then the command exits with a note. If
  you have access, `hf auth login` first.

## Picking a model

`mira-mini play` chooses weights for your machine: **CUDA gets the 1B**, **Apple silicon
gets the 364M laptop tier** (an MLX transformer + Core ML decoder, ~8 fps on a 2021 M1 Pro).
Override it:

```
mira-mini play --model 1b     # the 1B single-player model (wants a real GPU)
mira-mini play --model 364m   # the laptop tier, anywhere
```

## Options

| flag / env | effect |
|---|---|
| `--model {auto,1b,364m}` | which weights to run (default: auto, by device) |
| `--steps N` | sampler steps; 2 is the steadier default, 1 is smoother but drifts more |
| `--port N` | web UI port (default 8770) |
| `--no-browser` | don't open the browser automatically |
| `MIRA_HF_REPO` | use a custom Hugging Face weights repo |
| `MIRA_HOME` | where bundles are cached (default `~/.cache/alakazam-mira`) |
| `MIRA_DEVICE` | force `cuda` / `mps` / `cpu` |

## Weights and license

Model weights live on Hugging Face under
[alakazamworld](https://huggingface.co/alakazamworld) and are **CC BY-NC-SA 4.0**,
inherited from the training dataset
([kyutai/rocket-science](https://huggingface.co/datasets/kyutai/rocket-science), Rocket
League content used with Epic Games' permission). Non-commercial, share-alike, with
attribution. The model simulates one game and is a research demonstration, not a game
engine: long rollouts drift from exact physics.

## Credits

The architecture, training recipe, and dataset are General Intuition's and Kyutai's,
released openly with Epic Games ([mira-wm/mira](https://github.com/mira-wm/mira)).
MIRA Mini is Alakazam's independent reproduction and compression of that work.
