news

Tencent's Hy-Nomas report arrives days after the weights, and its abstract still carries no scores

The Apache 2.0 world model shipped with a 280-case benchmark Tencent built itself, and the October 27 technical report claims superiority in human evaluation without putting a number in the abstract.

By Daniel Ochoa · 29 October 2026

Tencent Hunyuan put an interactive video world model on Hugging Face under Apache 2.0, bundled a 280-case evaluation suite with it, and published no results for that suite. The technical report for Hy-Nomas, submitted to arXiv on October 27, 2026 by 38 authors, closes part of that gap but not the part that matters most: its abstract claims the model shows superior performance over existing world models in human evaluation, and names no score, no baseline, and no margin.

What actually shipped under Apache 2.0

Hy-Nomas is an 8 billion parameter segment-wise autoregressive diffusion transformer paired with a Qwen text encoder, camera-conditioning modules, and a PixelUMM video decoder, with a persistent key/value cache carrying context from one segment to the next. The distilled inference configuration uses 8 denoising steps and outputs 720 by 1280 at 24 FPS across 54 dual-stream MMDiT blocks. Users steer generation as it runs, introducing objects, changing scenes or styles, describing actions, or taking the camera somewhere else.

The hardware story has two versions. Tencent says a low-memory configuration runs on a single RTX 4090 using INT8 weights, an FP8 decoder and CPU offload, while the standard setup uses eight GPUs. The 24 FPS figure describes playback, not generation speed, which depends on hardware, resolution, segment length and history.

The benchmark comes from the same team as the model

Hy-Nomas-Bench covers four suites: multi-turn free-form text interaction, camera navigation, interleaved text and camera control, and spatial memory, meaning whether a scene still looks the same once the camera comes back to it. The release provides structured prompts, interaction sequences, keyboard timelines and reference camera trajectories. It is a real harness, and it is also the vendor's own, which means a strong result on it is a self-report until someone outside Tencent runs it.

Camera control and shot-to-shot consistency are the two dimensions that separate world models from clip generators, and both are scored separately in our benchmark rather than folded into an overall figure.

The Genie 4 comparison is made in a comment, not in a table

On the paper's Hugging Face page, the team positions the release directly against Google DeepMind's latest world model, writing that Hy-Nomas supports the "exactly same features as Genie4" while being fully open sourced. The feature overlap is arguable on paper. The quality comparison is not, because neither side has published numbers that the other could be measured against. Genie 4 arrived with trusted-tester access and no published benchmark results, and Hy-Nomas arrived with a benchmark and no published results on it.

The practical difference is that one of these you can download. Weights, inference code, a LoRA fine-tuning pipeline and the benchmark are all there under a permissive license, so the scores Tencent has not published are scores a third party can produce. That is a better position than the alternative, and it is still not the same thing as an independently verified claim.

Sources: Tencent on Hugging Face, arXiv, Hugging Face paper page, VP Land.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.