alakazamworld/mira-mini

Model Card for MIRA Mini

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 optimization of that work.

A 1B-parameter action-conditioned world model of Rocket League that generates gameplay one frame at a time, in real time, on a single GPU. It is the bundle mira-mini play downloads on CUDA machines (Apple silicon and CPU default to the 364M laptop tier; force either with --model 1b or --model 364m): our reproduction of the MIRA recipe at 1B (five times smaller than the original 5B), with its codec. The rest of the family, including a 364M student that runs locally on a 2021 MacBook, is listed under Model family below.

Model Details

Model Description

MIRA was released July 6, 2026 by General Intuition and Kyutai, in collaboration with Epic Games: code, data, and a published training recipe for a 5B-parameter action-conditioned world model of Rocket League. MIRA Mini is our independent reproduction of that recipe, trained in the seven days after the release. We trained the full stack: the codec, the world model, and the four-player multiplayer fine-tune.

We reproduced the 5B model at 1B parameters, five times smaller, then distilled it further. The consumer-hardware tier is a 364M student, more than ten times smaller than the original, small enough to run locally on a four-year-old MacBook through a parity-gated MLX port.

Reproduction was the first half of the work. The second half made it fast: a FlashDreams CUDA-graph serving port (2.75x over the reference implementation, bit-exact), a post-hoc few-step distillation that warm-starts from the base checkpoint, and a retrained small decoder with 8x fewer decode FLOPs. The 1B plays in a browser from a single L4; the 364M student plays on a laptop.

• Developed by: Alakazam
• Model type: Action-conditioned multiplayer world model (interactive video generation)
• 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 purposes only. The training dataset (kyutai/rocket-science) is licensed 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
• Training fork: Alakazam-studios/mira
• Technical report: alakazam.gg/mira-mini
• Hosted demo: play.alakazam.gg
• Upstream release: mira-wm/mira

Model family

One repository per runnable bundle, because mira-mini play downloads a whole repo and you should only pull the weights you run. This card describes the first row.

Repository	What it is	Runs on	Status
alakazamworld/mira-mini (this repo)	1B single-player world model, 8-step, plus the RAEv2 codec. 5x smaller than the 5B original	one L4/L40S/B200, or a 2021 MacBook via MLX	live
alakazamworld/mira-mini-364m	364M student with the small decoder, more than 10x smaller than the original. The local and laptop tier	modest GPUs and laptops	live
alakazamworld/mira-mini-psd	Same 1B model, few-step distilled (2 sampler steps instead of 8)	same GPU, ~1.3x delivered fps (19.4 measured on L40S)	live
alakazamworld/mira-mini-4p	Four-player multiplayer fine-tune (90k release; evals pre-registered at a 63k freeze)	one GPU per room	live

The local story runs on the same architecture at smaller sizes: the 1B through the MLX port on Apple silicon, and the 364M student for laptops and modest GPUs where you want real frame rates.

Uses

Direct Use

An interactive world simulator to play, probe, and measure: it takes the last 80 frames plus per-player controller actions and predicts the next frame, in a loop, in real time. It is also a working reference for anyone reproducing MIRA or studying what world-model training costs at small scale.

Out-of-Scope Use

Anything commercial. The license is non-commercial and share-alike, and commercial use of Rocket League content needs Epic Games. The model simulates one game and does not generalize beyond it; it is not a general video generator, a game engine, or a physics oracle.

How to Get Started with the Model

pip install alakazam-mira-mini
mira-mini play

One command starts the model, the room relay, and the web UI on your machine. Device auto-detection: CUDA, then Apple silicon (MPS), then CPU. The bundle this repository serves: world_model_config.yaml, checkpoint-<step>/checkpoint.pth, codec/, context/default.npz.

Training Details

Training Data

kyutai/rocket-science: Rocket League gameplay collected with publicly available bots, released by Kyutai under CC BY-NC-SA 4.0 with Epic Games' permission. We trained on the published dataset as released, no additions.

Training Procedure

The published MIRA recipe, run on rented spot GPUs over seven days. Components:

Component	Details
Codec	RAEv2: frozen DINOv3-L encoder, ViT-XL 1152x28 decoder, retrained from scratch (125k steps)
World model	1B-parameter DiT, 80-frame context, action-conditioned per player
Fine-tune	Four-player multiplayer, warm-started from the single-player model
Runtime	FlashDreams CUDA-graph serving (bit-exact) plus a parity-gated MLX port (matched to torch at display precision)

Deviations from the paper, all documented in the technical report:

• Post-hoc PSD warm-start. The paper also distills its pretrained model for few-step sampling, but leaves unspecified how the step-size (Δ) conditioning pathway is introduced into a checkpoint trained without one, nor the finetune recipe. We contribute that construction: inject the pathway with a zero-init output projection so step 0 is bit-identical to the base model, then distill at the recipe we found (lr 3e-5, warmup 100, PSD-only activation checkpointing; the from-scratch lr diverges). Two sampler steps replace eight.
• Decoder-only retrain (576x14): 8x fewer decode FLOPs, latents bit-identical by construction, validated by gated parity tests. Final decoder (60k steps): reconstruction LPIPS 0.18.
• 364M student (teacher-in-the-loop): 3.2x fewer DiT FLOPs, aimed at the consumer-hardware tier. Stage-1 distillation (40k) + stage-2 PSD (10k) complete; runs the full game live on a 2021 M1 Pro (below).

Evaluation

Measured on our training runs, July 2026, against the numbers reported in the MIRA paper. We report both columns because the honest comparison is the product.

Metric	MIRA Mini	MIRA (paper)	Note
Codec PSNR	28.6	29.7	from-scratch codec, 125k steps
Single-player gFID	12.8	10.7	ours at 52% of the paper's step budget
Single-player gFDD	0.45	0.55	lower is better
Multiplayer gFID	pending final eval	9.4 to 9.9	four-player fine-tune

Performance

Hardware	1B @ 8 steps	2-step (PSD) + small decoder
B200	25.7 fps	~40 to 50 fps (est)
L40S	14.5 fps	~35 fps, or 2 sessions per GPU (est)
L4	5.4 fps	~20 to 25 fps (est)
M1 Pro (2021 MacBook, MLX)	~1.4 fps	364M student: ~8 fps (2-step), ~12 fps (1-step)

The 364M student + 60k decoder run the full game live on a 2021 M1 Pro (MLX DiT + Core ML decoder). Two sampler settings are offered: 2 steps (the stable default, ~8 fps) and 1 step ("smooth mode", ~12 fps, 97.5% of 8-step sharpness; smoother to play but wobbles more per step). Measured numbers pass a parity gate: bit-exact for the CUDA/FlashDreams path, matched to torch at display precision for the MLX path; estimates are marked. Full deployment matrix in the technical report.

Bias, Risks, and Limitations

• Trained exclusively on Rocket League play collected with bots; it does not generalize beyond that world, and play styles beyond the bots' are underrepresented.
• Long rollouts lose texture fidelity. The model dreams plausible continuations, not exact physics.
• Non-commercial license, inherited from the dataset.

Credits

The architecture, training recipe, and dataset are General Intuition's and Kyutai's, released openly, in collaboration with Epic Games. MIRA Mini is Alakazam's independent reproduction and optimization of that work.

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)