seedleap/zing-0.5

TL;DR: Zing-0.5 is a 5B-parameter causal video world model with released weights (Apache 2.0) that generates continuous rollouts you can steer live — text prompts can change mid-generation and W/A/S/D/I/J/K/L keyboard actions alter camera and motion. It gets interactivity through the now-familiar recipe of causal KV caching plus few-step distillation (4-step DMD), running on a single H100. Weights and inference code are out; the technical report is not, and there are no benchmark numbers, so treat capability claims as unverified until you run it.

What it is

Most video generators are bidirectional-in-time: they denoise a whole clip at once, so you can't intervene once generation starts. Zing-0.5 belongs to the newer family of causal video diffusion models (the lineage of CausVid, Self-Forcing, and game-world models like Oasis/GameNGen) where frames are generated autoregressively and conditioning can change at any step. The distinguishing pitch here is the combination: text-initialized or single-image-initialized rollouts, prompt switching during generation, and continuous 8-key action control, all in one open 5B checkpoint on one GPU.

The pretrained/ directory (text encoder, tokenizer, VAE, kept separate from a 5B generator) strongly suggests this is built on an existing DiT video backbone — the 5B size and diffusers tagging are consistent with a Wan-style base — with causal attention and action conditioning trained on top. The README doesn't say, and there's no training-data or architecture detail; that's what the "coming soon" report needs to fill in.

How it works (from what's disclosed)

Three mechanisms make it interactive rather than batch:

1. Causal attention with KV caching. Each new latent frame attends only to the past, so generated history is cached and reused instead of recomputed — the standard trick that makes autoregressive rollouts affordable.
2. Attention sinks + sliding window. The local_attn_size / sink_size knobs (97/9 default, 33/5 for <80 GB) are StreamingLLM-style: keep a few "sink" tokens from the start of the sequence plus a local window, discard the rest. This bounds memory for long horizons at the cost of forgetting distant context — likely a contributor to the drift they admit to. Full-history attention is available via --local-attn-size -1.
3. Four-step DMD sampling. Distribution matching distillation compresses the denoising schedule to 4 steps per frame chunk, which is what makes "real-time" plausible on a single H100.

Actions and prompt edits are injected as conditioning at each generation step, so the model behaves like a controllable simulator rather than a clip generator.

What's actually there — and what isn't

There: weights on HF (generator model.pt state dict, strictly loaded, plus text encoder/tokenizer/VAE), a standalone inference repo (seedleap/zing-world-model) with ready-to-run Action T2V and Action TI2V JSONL examples including reference images, memory configs for ≥80 GB and smaller GPUs, and 24 FPS H.264 output.

Not there: the technical report, any quantitative results (FVD, latency, actual achieved FPS — "real-time" is asserted, not measured), training details, comparisons to Matrix-Game / Oasis / Self-Forcing-style baselines, and any interactive demo — the released harness is offline JSONL-in, MP4-out, so you're scripting action sequences, not driving with a keyboard yet. Downloads are at zero and there's no community signal. The README is honest about limitations: visual drift on long rollouts and action responsiveness that varies with scene content and viewpoint.

Bottom line

The capability class — open weights for a promptable, action-conditioned, streaming world model at 5B — is exactly the interesting frontier, and the sink/window memory design is a sensible answer to long-horizon cost. But this is a day-zero release from an unknown team with no paper and no numbers. Worth an hour on an H100 to check two things yourself: how fast the 4-step rollout actually is, and how quickly identity and physics drift once the sliding window starts evicting history.

Try it

git clone https://github.com/seedleap/zing-world-model.git
cd zing-world-model

ZING_MODEL=/path/to/Zing-0.5   # download from hf.co/seedleap/zing-0.5
CUDA_VISIBLE_DEVICES=0 bash run.sh \
  --pretrained-dir "$ZING_MODEL/pretrained" \
  --checkpoint "$ZING_MODEL/generator/model.pt" \
  --messages examples/case3_action_t2v.jsonl \
  --output-dir outputs/case3 --seed 0
# <80 GB VRAM: add --local-attn-size 33 --sink-size 5

Model card: https://huggingface.co/seedleap/zing-0.5