Skywork releases Matrix-Game 4.0 weights and Apache-2.0 inference code

On 15 October 2026, the Matrix-Game repository news line says Skywork AI released Matrix-Game 4.0. Inference, a browser demo and game source are in that repo. Weights and game assets for Medieval Castle, Ancient Egypt and City Racing are at Skywork/Matrix-Game-4.0. Pandaily, with a CMS publishedDate of 2026-10-15T03:40:00.000Z, describes Skywork AI as the AI unit of Kunlun Tech. The Hugging Face card and the GitHub pages read here say Skywork AI, and they do not name Kunlun.

The repository README says an executable engine handles gameplay, physics and player input, and a causal video model turns low-poly RGB frames into detailed visuals. The model card says the same split, and it adds that HUD and sound stay outside the model as frame-aligned layers. The GitHub README calls Medieval Castle and Ancient Egypt third-person shooting scenes, and City Racing a driving game. On the Hub those checkpoints sit in fps-castle, fps-egypt and driving_game. The card’s base_model field is alibaba-pai/Wan2.2-Fun-5B-Control. It says each scene checkpoint has 5,001,016,512 BF16 parameters, that the weights need the Matrix-Game 4.0 inference source, and that they are not a drop-in Diffusers pipeline. Training code and training data are not included.

The project page describes a harness: the engine keeps state and draws a low-poly proxy, and a few-step causal video model renders each chunk. Its paper and code buttons still read “soon” on the page opened 15 October, while the GitHub news line of the same morning says the code is released. The page says the video model reads the proxy, a dynamic caption and reference frames, and generates one latent, four frames, per step, with a causal DiT distilled to one or two denoising steps.

Two licenses, plus an older MIT file

The Hub LICENSE is the Apache License, Version 2.0. The card says project scene weights, sink images and game assets are Apache-2.0 from Skywork AI, that shared Wan and umT5 resources keep their upstream Apache-2.0 terms, that TAEHV by Ollin Boer Bohan is MIT, and that racing audio includes CC0 and CC BY 3.0 material. Those upstream files were not re-opened here. The card is the source for that sentence.

The v4 code LICENSE is also Apache License 2.0. A NOTICE file in that directory says “Copyright 2026 Skywork AI and contributors.” The repository root LICENSE is the MIT License, copyright 2025 SkyworkAI and contributors. The root README says Matrix-Game 4.0 is Apache-2.0 and that earlier versions keep their existing terms, including that MIT file. GitHub’s API reports the repository license as MIT, which matches the root file and not the v4 file.

What the speed pages actually say

The model card says there is no guaranteed frame rate, no guaranteed consumer-GPU memory fit, and no exact scene reconstruction, and that the card does not claim a performance benchmark. Output resolution on the card is 480 by 864 RGB. A quality preset uses the Wan VAE and two denoising steps, 1000 and 500. A realtime preset uses TAEHV and one step, 1000.

The v4 README says model inference is tested on an A100 80 GB with PyTorch 2.12.0 and CUDA 13.2. A note dated 2026-10-14 says optimized inference code for GeForce RTX 50-series GPUs was planned for later that week. The same README includes an rtx5090 preset that replaces the DiT runtime with FP8 linears, fused kernels and CUDA graphs. Its measurement table is on one NVIDIA A100-SXM4-80GB, at 864 by 480, one denoising step, TAEHV, excluding startup and warmup. Castle and Ancient Egypt, fixed prompts, are listed at about 38 to 39 frames per second of model processing in an online session. Racing is about 38.5 frames per second of server-side pipeline throughput with fixed prompts, and about 28.9 with dynamic prompts. The README says those figures are not browser presentation frame rate and not action-to-photon latency.

The project page prints five headline figures: 83 FPS for 1-step denoising on an RTX 5090, 112 ms from a key press to the first reacting frame, 48 ms to denoise one chunk (one latent frame, four frames), 480p streaming output, and one GPU holding the text encoder, DiT and VAE. Those are the page’s figures. Pandaily, citing QbitAI, also reports about 57 frames per second with two steps and about 18.5 GiB of video memory, and it says those figures come from the company. QbitAI was not opened. The 57 frame and 18.5 GiB figures were not on the project-page stats or the A100 table in the v4 README.