alakazamworld/mira-mini-4p

Model Card for MIRA Mini 4P

The four-player multiplayer tier of MIRA Mini: a 1B action-conditioned world model that simulates a 2v2 Rocket League match for four players at once, each with their own first-person view, conditioned jointly on all four action streams. It is our reproduction of the MIRA multiplayer recipe at 1B scale (the original demo is 5B).

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.

Model Details

Model Description

The multiplayer wrapper tiles the four players' views into one joint latent frame (stacked along height) and runs a single diffusion transformer over the tile, so all four views are predicted jointly and stay mutually consistent: one ball, one match, four synchronized dreams of it. Per-player action streams are embedded, combined, and condition every step. Training uses the paper's per-player action dropout, so any seat can be handed to the model instead of a controller: the model imagines what that player would have done (the autopilot mode in the demo).

Following the paper's budget-split finding, the model is warm-started from our single-player checkpoint (52k steps) and fine-tuned on four-view multiplayer data. Two inference-time options ship with the serving stack, both documented in the technical report: an action-guidance mode that amplifies a player's steering authority (useful at this model's training budget, where early-window controllability is still maturing), and a distilled small decoder that renders the same latents about 8x cheaper.

• 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 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
• Base 1B model: alakazamworld/mira-mini
• Hosted demo: play.alakazam.gg
• Upstream release: mira-wm/mira

How to Get Started

pip install alakazam-mira-mini
mira-mini play --model mira-mini-4p

That starts the model, a room relay, and a web UI; share the room code and up to four people play in the same simulated match (one GPU per room). Seats without a human are driven by the model (autopilot). The bundle is world_model_config.yaml, checkpoint-<step>/checkpoint.pth, codec/, context/default.npz.

Training Details

Training Data

kyutai/rocket-science: four synchronized first-person recordings per match with per-player action streams, 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

The paper's two-stage recipe: single-player pretraining (52k steps), then the multiplayer fine-tune on 4 nodes x 8 H100 (preemptible), global batch 32 tiled frames (128 player-views per step), per-player action dropout p=0.1, checkpoint every 1,000 steps with a fleet-consistent resume protocol that survived four spot preemptions with zero lost steps. Evaluations were pre-registered at a 63k freeze; the released checkpoint is 90k multiplayer steps, requalified on the full suite before shipping (the paper's demo model trains 100k on ~5x the data at 5B parameters; treat this tier accordingly).

Evaluation

Metric	MIRA Mini 4P	MIRA (paper)	Note
MP val loss	monotone through the 90k release (0.320 at 80k, last validation)	n/a	no overfit signature
MP gFID (4s, 10-step paper protocol)	24.1 at release; 24.7 at the 63k freeze (27.3 at 50k, improving)	9.4 to 9.9	paper is 5B / 100k steps / ~5x data; paired-Frechet, floor ~6; at the 8-step serving setting (freeze): 28.7/29.0
Long-rollout stability	10 min, zero freezes; 30s park-OOD melts 4.1x at the freeze, snaps back in 10s; at release the same park produces no visual melt (drift <=1.6x baseline) and the largest excursions (2.7x) are hallucinated goal/replay broadcast events, frames coherent	"stable indefinitely"	first 10-minute rollouts on this tier; evidence clips in the repo
Controllability	measured per-checkpoint	ARR (Fig. 13 of the paper)	our seed-controlled divergence ladder; full curves in the report

Controllability at this budget is the honest caveat: steering is measurable and improves with training (the report publishes the per-checkpoint curve; early-window authority 0.281 un-guided at the 90k release vs 0.261 at the freeze, and 0.355 at the demo's default guidance w=4), and the serving stack's action-guidance option amplifies it at inference; boost and jump respond less than steering, matching the paper's finding that rare actions are recovered last.

Performance

Runtime	fps (B200, 8 steps, four views)
Reference implementation	~8.6
FlashDreams CUDA-graph port (bit-exact)	~10.8
+ distilled small decoder	higher; measured numbers in the technical report

Bias, Risks, and Limitations

• Bot-collected Rocket League only; no transfer beyond the game.
• Early-window steering authority is below the single-player model's at this training budget (quantified in the report); action guidance narrows the gap at ~2x inference cost and is applied per view (the commanded car's view only — a scalar guidance weight amplifies the driver's actions onto every view).
• Action attribution is a trained property, incomplete at this budget: all four players' actions are mean-pooled into one conditioning stream, so commanded actions bleed weakly into other views as a global motion prior (measured in the report; sharpens with training).
• Replays after goals are hallucinated (the context window is ~4 s, matching the paper's own limitation), and out-of-distribution play can degrade the image before it recovers.
• Non-commercial license, inherited from the dataset.

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)