LynnReal-Hand streams a hand-controlled world model at 18 FPS
The open model adds continuous hand motion on top of text and camera control, and its authors report 93.02 MANO Dice on a benchmark no third party has run.

Key takeaways
- LynnReal-Hand is an open, hand-controlled video world model from the LynnReal AI group; the paper is a technical report dated November 17, 2026, and the inference code and weights are released on GitHub and Hugging Face.
- In the paper's own HandBench results, the full-sequence model scores 181.67 FVD and 93.02 MANO Dice, while the three-step causal Fast model holds 92.49 MANO Dice.
- The authors report 540p streaming at 18.03 FPS with 741 milliseconds of latency and 34.89 GiB peak GPU memory on a single H200.
- The paper says it will release its 100K-clip dataset and the HandBench benchmark, but no download link for either appeared in the project's repository or model cards as of November 19, 2026.
The LynnReal AI group released LynnReal-Hand, a video world model that adds continuous hand motion as a control channel on top of the text and camera controls other world models already offer, in a technical report dated November 17, 2026. The paper describes hands as the interface that can communicate "intentions that are difficult to specify through text or navigation commands." Every number below is the authors' own, measured on their own benchmark, and no third party has run it.
What hand control adds. Interactivity in this field has centered on navigation, camera motion or high-level instructions, which the paper says "determine where users look or move, but offer limited ability to manipulate the objects they encounter." LynnReal-Hand conditions generation on articulated hand motion, taken from per-frame MANO hand meshes rasterized into the frame (red for the right hand, blue for the left), together with camera poses. The paper says that lets a user reach, grasp, release and move both hands through a scene.
Why the data is the bottleneck. The authors call large-scale annotated interaction data scarce, and note that hands and their contact points fill only a small part of each frame, so a standard global training goal under-supervises the fingers. They assemble 100K interaction clips across real and synthetic scenes and add hand-aware supervision, a hand-region latent loss plus LPIPS and DINOv2 scores computed inside hand crops, to force the model to learn the small regions. The pipeline then runs full-sequence pretraining, a causal flow map for three-step generation, region-aware score distillation, and a compact Ultra-Fast-VAE decoder (361.2 million parameters) to cut decoding time.
The numbers, on the authors' own HandBench. That benchmark is 200 sequences: 160 in-domain clips drawn from eight sources, and 40 held-out clips from H2O and TACO for out-of-distribution testing. The full-sequence model, LynnReal-Hand-Base, scores 181.67 FVD and 93.02 MANO Dice. The three-step causal model, LynnReal-Hand-Fast, holds 92.49 MANO Dice at 237.33 FVD. Against the strongest fine-tuned baseline the paper ran, Hand2World in its bidirectional form, the paper reports Base cutting FVD by 38.3% (181.67 against 294.32) and lifting MANO Dice from 90.60 to 93.02.
For streaming, the authors report 540p output at 18.03 FPS with 0.741 seconds of latency and 34.89 GiB of peak GPU memory on a single H200.
What is released. Code and weights are out: the GitHub repository carries the inference code and example inputs, and three checkpoints sit on Hugging Face (Base, Fast-preview and Ultra-Fast-VAE). The 100K dataset and HandBench are promised in the paper but no download link for either appeared as of November 19, 2026. The weights ship under the MiniMax H3 Community License Agreement and need one H200-class GPU with FlashAttention-3.
The paper's own limitations section is worth reading before the optimism. The model's local two-chunk causal window keeps streaming fast but carries no long-term 3D memory, so the authors say object permanence could suffer when an object leaves and re-enters view. The model card adds that generated interactions "can exhibit hand artifacts, object drift, or temporal instability."
For a creator, this is a research release, not a tool: a 33-billion-parameter model that needs a data-center GPU. If the streaming figure holds up in independent hands and the weights stay open, the interesting part is the control channel rather than the frame rate, because a scene you can grab by hand is a different kind of preview than one you can only prompt.
Code and weights are at the LynnReal repository; the paper is on arXiv.