alakazamworld/mira-mini-psd

Model Card for MIRA Mini PSD

The few-step variant of MIRA Mini: the same 1B action-conditioned world model of Rocket League, distilled so that two sampling steps replace eight. The sampler does 4x less work per frame; delivered frame rate on the same GPU is about 1.3x (14.5 to 19.4 fps measured on L40S) because decode still dominates, until the small decoder is swapped in. Visual quality matched the 8-step baseline on our evaluations.

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

Identical architecture to alakazamworld/mira-mini (a 1B diffusion transformer in the latent space of a representation-autoencoder codec), plus the paper's progressive self-distillation (PSD), applied post hoc. Sampling a latent frame integrates the flow-matching field; PSD distills the model so one large integration step replaces two smaller ones, which lets two steps reach the quality that previously took eight.

The paper distills its pretrained model and leaves unspecified how the step-size (Δ) conditioning pathway is introduced into a checkpoint trained without one. This model uses our construction: the pathway is injected with a zero-initialized output projection, so at step 0 the model is identical to the base checkpoint, and the distillation signal fades in from there (10k steps, learning rate 3e-5, warmup 100; the from-scratch learning rate diverges on a converged checkpoint).

• Developed by: Alakazam
• Model type: Action-conditioned 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
• Upstream release: mira-wm/mira

How to Get Started

pip install alakazam-mira-mini
mira-mini play --model mira-mini-psd

The bundle is world_model_config.yaml, checkpoint-10000/checkpoint.pth, codec/, context/default.npz. Two steps is the intended setting; 4 and 8 also work (quality is flat across them after distillation, so extra steps mostly cost frame rate).

Training Details

Post-hoc PSD fine-tune of the mira-mini 1B checkpoint: 10k steps on 8 preemptible H100s, PSD loss applied stochastically on 10% of updates (the paper's mixing rate), ground-truth flow term otherwise. Validation loss improved from 0.3952 (base, at handoff) to 0.3741, and the 2-step teacher-forced LPIPS ends below the base model's 2-step LPIPS with the margin still growing at 10k (−0.0021 on our held-out shard).

Training data (through the base model): kyutai/rocket-science, as released, no additions.

Performance

Hardware	8 steps (base)	2 steps (this model)
L40S	14.5 fps	19.4 fps measured (live serving probe)
B200	25.7 fps	not benched at 2 steps
L4	5.4 fps	not benched at 2 steps

Measured numbers come from serving probes on the stated hardware; estimates are marked. The serving runtime is the FlashDreams CUDA-graph port (bit-exact against the reference implementation at matched settings).

Bias, Risks, and Limitations

• Same limitations as the base model: bot-collected Rocket League only, no transfer, plausible continuations rather than exact physics.
• Few-step sampling drifts slightly more per step than 8-step sampling on long rollouts; quantified in the report's physics-verification section: in ten-minute rollouts the distilled rungs ride an elevated-but-flat distance band (a level shift, not progressive melting), and per-action recoverability degrades tail-first (rare actions like air-rolls before common ones) — aggregate numbers hide exactly that.
• 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)