Luffuly/streaming-code-world-model

Streaming Code World Model

LoRA checkpoints for Streaming Code World Model: causal, chunk-by-chunk video generation from a live proxy, on top of the released Code World Model (MiniMax-H3 Ref2VA + CWM LoRA). Both files are LoRAs for the released MiniMax-H3 model; install the CWM release first (see the GitHub README).

file	model	sampling	CARLA MSE ↓	s / chunk (H200)
streaming_df.safetensors	streaming diffusion forcing	--steps 35	2231	47.9
streaming_dmd_4step.safetensors	4-step DMD, the streaming model as teacher	--steps 4 --rule x0 --shift 5 --kv-reuse	1660	5.1

hf download Luffuly/streaming-code-world-model --local-dir ckpt
EVAL_ARGS="--steps 4 --rule x0 --shift 5 --kv-reuse" \
    sbatch scripts/eval/eval.sbatch ckpt/streaming_dmd_4step.safetensors out/ clip_0048

Inference runs on one NVIDIA H200 (141 GB); peak memory 123 GB (bf16), 108 GB (--fp8).