LongCat-Video Technical Report

TL;DR

LongCat-Video is Meituan’s 13.6B-parameter dense foundation video model — MIT-licensed open weights — that unifies Text-to-Video, Image-to-Video, and Video-Continuation tasks under a single backbone. The headline pitch is native minute-long generation: the model is pretrained on Video-Continuation (not just T2V), so long-form rollouts come out of training rather than being bolted on at inference. Efficiency comes from a coarse-to-fine temporal/spatial generation strategy plus Block Sparse Attention; quality comes from multi-reward GRPO RLHF. In Meituan’s own MOS evaluation on internal prompts the model trades blows with Veo 3, PixVerse-V5, Wan 2.2-T2V-A14B (T2V) and Seedance 1.0, Hailuo-02, Wan 2.2-I2V-A14B (I2V), reaching parity-class Overall Quality at roughly half the active-parameter count of Wan 2.2-A14B.

Key claims

13.6B-parameter dense video DiT, open-sourced under MIT — released alongside two avatar-tuned variants (LongCat-Video-Avatar, LongCat-Video-Avatar-1.5) with Whisper-Large-v3 audio encoding and 8-step distillation in v1.5 [README §Latest News, §Model Download].
One model natively supports Text-to-Video, Image-to-Video, and Video-Continuation tasks within a single framework rather than three task-specific checkpoints [README §Key Features].
Pretraining includes Video-Continuation as a native task, which the authors claim is what enables minute-long generation without color drift or quality degradation [README §Key Features].
720p / 30fps generation in “minutes” via a coarse-to-fine generation strategy along both temporal and spatial axes; Block Sparse Attention is named as the further efficiency lever at high resolution [README §Key Features].
Post-trained with multi-reward Group Relative Policy Optimization (GRPO) — the same GRPO-on-video-diffusion recipe lineage as DanceGRPO and HuDA, here applied to a foundation-scale dense model [README §Key Features].
T2V MOS on Meituan’s internal benchmark vs frontier proprietary + open baselines (Veo 3 / PixVerse-V5 / Wan 2.2-T2V-A14B): Text-Alignment 3.76, Visual Quality 3.25, Motion Quality 3.74, Overall Quality 3.38 — Overall Quality between Wan 2.2-A14B (3.35) and Veo 3 (3.48), at 13.6B dense vs 28B-total / 14B-active MoE [README §Text-to-Video Table].
I2V MOS vs Seedance 1.0 / Hailuo-02 / Wan 2.2-I2V-A14B: Image-Alignment 4.04, Text-Alignment 3.49, Visual Quality 3.27 (best of the four), Motion Quality 3.59, Overall Quality 3.17 (vs Seedance 1.0 3.35) [README §Image-to-Video Table].
FlashAttention-2 is the default attention backend in the shipped model config; FlashAttention-3 and xformers are drop-in swaps via the model config [README §Installation].
Reference inference supports context-parallel multi-GPU rollout out of the box (--context_parallel_size=N), and a --enable_compile flag for torch.compile-friendly serving across all four entry-points (T2V, I2V, Continuation, Long-Video) [README §Run Text-to-Video, §Run Long-Video Generation].

Method

LongCat-Video is a 13.6B-parameter dense video DiT trained with three coupled task heads — Text-to-Video, Image-to-Video, and Video-Continuation — sharing the same backbone. The architectural / training details that matter mechanically for long-form output are (a) Video-Continuation as a first-class pretraining task rather than an inference-time scaffold, so the model directly learns to condition on a prefix of generated frames; (b) a coarse-to-fine generation strategy along both temporal and spatial axes, in the same family as Wan 2.2’s hierarchical sampling and SANA-Video’s block-linear sampling — generate a low-rate / low-resolution draft, then refine; and (c) Block Sparse Attention to keep the per-step attention cost bounded as the context window stretches to multiple seconds of latent video.

Post-training is a multi-reward GRPO stage — group-relative advantage over rollouts scored against a basket of reward models, with the per-reward weighting acting as the implicit “what does quality mean” knob. The full architectural diagram, sparse-attention block layout, reward composition, and exact RLHF training schedule are in the technical report (arXiv:2510.22200); the README confirms only that the recipe is GRPO-family and that the rewards are multi-objective.

Results

Meituan’s own MOS evaluation on an internal benchmark is the headline number. T2V Overall Quality (out of 5) lands LongCat-Video at 3.38, between Wan 2.2-T2V-A14B (3.35, 28B-total / 14B-active MoE) and Veo 3 (3.48, proprietary), with Motion Quality 3.74 trailing Veo 3 (3.86) and matching Wan 2.2 (3.78) [README §Text-to-Video Table]. The parameter-count framing matters: LongCat-Video is 13.6B dense (every parameter active per token), versus Wan 2.2’s 14B active out of 28B total, so the comparison is roughly compute-matched per step.

On I2V vs Seedance 1.0 (proprietary) / Hailuo-02 (proprietary) / Wan 2.2-I2V-A14B, LongCat-Video wins on Visual Quality (3.27 vs Seedance’s 3.22) but trails on Image-Alignment, Text-Alignment, and Motion Quality, settling at Overall Quality 3.17 vs Seedance’s 3.35 [README §Image-to-Video Table]. No external public-benchmark numbers (VBench, MovieGenBench, PhysicsIQ) are reported in the README; the technical report (2510.22200) is where those live if at all.

Why it’s interesting

This is the third Meituan LongCat foundation release on the wiki — after LongCat-Flash Technical Report (LongCat-Flash LLM), LongCat-Flash-Omni Technical Report (LongCat-Flash-Omni), and LongCat-Image Technical Report (LongCat-Image) — and the first one explicitly framed as a world-model first step. The contrast with HunyuanVideo 1.5 Technical Report (HunyuanVideo 1.5, also Chinese-lab, also open) is on the long-form axis: Hunyuan pushes single-clip quality; LongCat-Video pushes native minute-long Video-Continuation pretraining. The GRPO-on-video-diffusion recipe slots into the Reasoning RL family alongside HuDA (Human detectors are surprisingly powerful reward models) and DanceGRPO: Unleashing GRPO on Visual Generation — a multi-reward composition is now the de facto post-training step for any open foundation video DiT. The Block Sparse Attention + coarse-to-fine strategy puts this in dialogue with Radial Attention: O(n log n) Sparse Attention with Energy Decay for Long Video Generation and SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer on the cost-vs-length Pareto, though the README doesn’t give enough detail to position the recipe head-to-head.