Tencent's Hy-Nomas World Model Is Apache-2.0 With 196 GB of Weights and No Published Scores
A.I. / news
Tencent's Hy-Nomas World Model Is Apache-2.0 With 196 GB of Weights and No Published Scores
The text-and-camera video world model runs on one RTX 4090 in an INT8 mode, but its card says the technical report is still coming and gives no parameter count.
Tencent has put Hy-Nomas, an interactive video world model, on Hugging Face under the Apache-2.0 licence, with about 196.4 GB of downloadable checkpoints and no benchmark results. The Hugging Face API record shows the repository was created on Oct. 21 and last modified on Oct. 25.
As of Tuesday the model page showed 86 likes and 155 downloads in the last month, according to the card. The GitHub repository showed 263 stars and 28 forks.
What the Hy-Nomas model card says
The model card describes a model that generates video worlds one segment at a time from text instructions and camera trajectories, across multiple turns. Output is 720 by 1280 at 24 frames per second, using 8 denoising steps in the distilled configuration. The GitHub README says 4 to 8.
The network has 54 double-stream blocks, a hidden size of 2,048 and 16 attention heads. The text encoder is Qwen2.5-VL-7B-Instruct. The card gives no total parameter count and no release date.
The card introduces Hy-Nomas-Bench, with four benchmarks and 280 test cases, and publishes no scores. It says a technical report is "Coming Soon." The GitHub release notes say the technical report was updated on Oct. 26, and the card's wording had not changed when it was read.
What it takes to run
| Setup | Hardware | Source |
|---|---|---|
| Standard inference | 8 GPUs in one group | Model card |
| Low-memory inference | One RTX 4090, INT8 weights, CPU offload | Model card |
| Interactive demo, tested | 8 B200 GPUs | GitHub README |
| LoRA fine-tuning example | 128 GPUs | GitHub README |
The manual install needs Python 3.12 or later and CUDA 13.0. A Docker image, hynomasoss/hynomasoss:cu130, is also listed. The single-GPU path caps active history at 18 latent frames and archives the rest in system memory, so RAM requirements grow with session length.
The repository holds distilled FP32 and INT8 checkpoints, a base autoregressive model, two LoRA adapters (cinematic and dual-arm robotics) and two decoder variants. The robotics adapter was trained on the RealSource-World dataset, the card says.
What is unclear
The card describes Nomas-PE, a prompt enhancer the web demo uses, as an online service. The repository nonetheless contains a PE folder of 18 weight shards. The pages read do not say whether that folder is the same model.
The AI-API vendor orcarouter wrote that its estimate of about 8.5 billion parameters comes from file sizes and that Tencent has not confirmed it. That blog also reported the technical report link returning a 404 at the time it was written. The same post gives the model repository's single commit as Oct. 23, two days after the creation date in the API record.
The card says third-party models and components keep their own licences, which matters for the Qwen text encoder. It lists no territorial limits on the Apache-2.0 grant.
GitHub's notes list community integrations added on Oct. 26 for SGLang, ComfyUI and Reactor Runtime, and WanGP v18.81 on Oct. 25. For how open-weight releases are being scored elsewhere, see our coverage of the DeepSeek V4.1-Flash scores, and for a local desktop app built on downloadable weights, the Hob release.
The next fixed point is Tencent's technical report, which the card still lists as coming.
Sources
More in A.I.
- 01Lean Creator Leo de Moura Says AI Makes Proofs Cheap. Here Is What the Checkers ShowOn a podcast released October 29, the AWS scientist who built Lean claimed AI-written proofs now run to 50 million lines. The claims check out only in part, and he names the two gaps himself.
- 02Claude's Watermark Quality Claim Rests on Single Responses, Blogger SaysAnthropic says watermarking does not affect output quality. A post citing Google's own SynthID-Text paper says the tests never measured what happens when you ask the same question twice.
- 03Nanbeige4.3-33B Ships Under Apache-2.0 With a 524,288-Token Window and Self-Reported WinsNanbeige Team's 33B model is built as a teacher for smaller agents. Its own table has Qwen3.8-27B ahead on SWE-Bench Pro and GDPval.
- 04GPT-5.6 Sol Launched a Factorio Rocket on Peaceful Mode, With Hauling Done by HandBlogger Ryan Madden's single run took 13.5 in-game hours. The agent carried supplies between areas because it never linked some production lines.