alakazamworld/mira-mini-364m

Model Card for MIRA Mini 364M

A 364M-parameter action-conditioned world model of Rocket League that generates gameplay one frame at a time, in real time, on a 2021 MacBook. It is the consumer-hardware tier of MIRA Mini, distilled from our 1B reproduction of the MIRA recipe.

Built on MIRA, released July 6, 2026 by General Intuition and Kyutai with Epic Games: code, dataset, and a published training recipe. MIRA Mini is Alakazam's independent reproduction and compression of that work.

Model Details

Model Description

The model takes a rolling window of the recent past, about four seconds of play, plus each player's controller inputs, and predicts the next latent frame, autoregressively. The codec runs at 10 Hz in latent space and video is reconstructed at 20 fps. It is a diffusion transformer operating in the latent space of a representation-autoencoder codec: a frozen DINOv3-L encoder, a linear bottleneck to 32 channels at one latent per 32x32 pixels, and a space-time ViT decoder.

This 364M student was distilled from our 1.18B reproduction, not trained from data. During distillation the 1.18B teacher runs alongside the student and the student regresses the teacher's velocity field, with a 0.25-weighted ground-truth flow-matching term as an anchor. A second stage adds the paper's progressive self-distillation (PSD) so the model reaches usable quality in two sampling steps instead of eight. The codec ships with a retrained decoder at 576 width by 14 depth (down from 1152 by 28), which cuts decode cost about 8x while leaving the latent space untouched, so it renders any world model trained on that codec.

• Developed by: Alakazam
• Model type: Action-conditioned world model (interactive video generation)
• Language(s): English (in-game UI text; the model is otherwise visual)
• License: CC BY-NC-SA 4.0, inherited from the training dataset
• Reproduction of: MIRA (General Intuition and Kyutai, with Epic Games)

This model is for demonstration and research only. The training dataset (kyutai/rocket-science) is CC BY-NC-SA 4.0, with Rocket League content used by Epic Games' permission. These weights inherit that license: non-commercial, share-alike, with attribution.

These weights are an independent release by Alakazam. They are not released by, associated with, or endorsed by General Intuition, Kyutai, or Epic Games.

Model Sources

• Player (one command): Alakazam-studios/alakazam-mira-mini
• Technical report: alakazam.gg/mira-mini
• 1B single-player model: alakazamworld/mira-mini
• Upstream release: mira-wm/mira

Uses

Direct Use

Play it locally, or read it as a worked example of compressing an interactive world model to consumer hardware. This is the bundle mira-mini play downloads on Apple silicon and CPU machines (CUDA machines default to the 1B; force either with --model 364m or --model 1b) to run the game on your own machine, offline. On Apple silicon it runs the transformer through an MLX port and the decoder through Core ML.

Out-of-Scope Use

Non-commercial only, and Rocket League content needs Epic Games for any commercial use. The model simulates one game and does not transfer beyond it. It is not a general video generator, a game engine, or a physics oracle: it predicts plausible next frames, and long rollouts drift away from exact physics.

How to Get Started

pip install alakazam-mira-mini
mira-mini play

That starts the model, the room relay, and a local web UI. Device selection is automatic: CUDA, then Apple silicon (MPS/MLX), then CPU. The bundle in this repository is world_model_config.yaml, checkpoint-10000/checkpoint.pth, codec/, context/default.npz, and decoder60k.mlpackage (the Core ML decoder for the Mac path).

Two sampler settings are exposed. Two steps is the default and the steadier picture; one step is a faster mode that is smoother to move in but drifts more per step.

Training Details

Training Data

kyutai/rocket-science: Rocket League play collected with publicly available bots, released by Kyutai under CC BY-NC-SA 4.0 with Epic Games' permission. About 15,800 matches, roughly 2,000 hours. We trained on the released dataset with no additions.

Training Procedure

Two distillation stages on eight H100 GPUs, rented preemptible.

• Stage 1, cross-scale distillation (40k steps). The 364M student regresses the 1.18B teacher's velocity predictions on shared draws, anchored by a ground-truth flow term at weight 0.25. The teacher is attached outside DDP, EMA, and the checkpoint, so a saved checkpoint holds only the student. Final ground-truth validation loss 0.425, against the teacher's own 0.374.
• Stage 2, progressive self-distillation (10k steps). The step-size conditioning pathway is added to the trained checkpoint with a zero-initialized output projection, so the model is identical to its stage-1 self at step 0, then distilled so two sampling steps match many-step quality. Learning rate 3e-5, warmup 100.
• Decoder (60k steps, separate run). Only the decoder is retrained, at 576 by 14, against the frozen codec encoder. Final reconstruction LPIPS 0.18.

The MLX transformer port matches the reference PyTorch model to the parity harness's print precision. The Core ML decoder matches PyTorch at 73 dB PSNR.

Compute

Preemptible H100 spot instances on Google Cloud. The student and decoder runs are each a single-node spot finetune; the few-step pilot validated its mechanism on one rented GPU-hour before any fleet spend.

Performance

Measured on an M1 Pro (2021 MacBook, 16 GB), running the transformer on MLX and the decoder on Core ML.

Setting	Live frame rate	Note
2 steps	~8 fps	default; steadier
1 step	~12 fps	faster; 97.5% of the 8-step sharpness, drifts more

The transformer forward is 58 ms; a two-step latent takes 168 ms, a one-step latent 112 ms. On stronger hardware (an M3/M4 laptop, or a recent discrete GPU) the same weights run two to three times faster. Full deployment numbers, including the Neural Engine investigation, are in the technical report.

Bias, Risks, and Limitations

• Trained only on bot-collected Rocket League play. It does not generalize past that game, and human play styles the bots did not produce are underrepresented.
• Long rollouts lose texture and drift from exact physics; the model dreams a plausible continuation, not a simulation.
• Non-commercial license, inherited from the dataset.
• Distilled from a 1B reproduction that itself trails the 5B original (codec PSNR 28.6 vs 29.7), so image quality is below the paper's, and below what the same recipe reaches at full scale.

Citation

Cite the MIRA paper and link this repository.

@article{hu2026mira,
  title  = {Multiplayer Interactive World Models with Representation Autoencoders},
  author = {Hu, Anthony and others},
  year   = {2026},
  note   = {arXiv:2607.05352}
}

Model Card Authors

Alakazam (alakazam.gg)