Tencent publishes Hy-Nomas technical report and 280-item benchmark

Researchers and rival labs can now test long-horizon consistency, instruction following and camera control against a published set instead of a promotional clip. The numbers are Tencent's own: its human evaluation puts Hy-Nomas above existing models, and the Genie 4 comparison footage was not produced under matched, independent conditions.

What Tencent actually published

The 30 October release is a report and an evaluation set, not a product launch. Tencent had already announced Hy-Nomas. The new material covers internal structure and evaluation method.

Hy-Nomas is an 8-billion-parameter video diffusion model. It takes free-form text instructions and continuous camera control across multiple turns. The weights are on Hugging Face under Apache 2.0 at around 196GB. Code is on GitHub, and a hosted demo is available. The Hugging Face release dates from late October.

How the model is built

Hy-Nomas is a segment-autoregressive diffusion transformer, not a single-shot video generator. It produces variable-length segments conditioned on a cached history. The world memory is a KV cache carried forward and refreshed as new segments are generated. Instructions apply only within a limited time window. An attention mechanism sets that window. Roughly 8 billion parameters sit across 54 dual-stream MMDiT blocks. The hidden size is 2,048. An optional prompt enhancer, Nomas-PE, rewrites loose prose into structured JSON captions. Output arrives at 720 by 1280 at 24 frames per second. A fine-tuned PixelUMM decoder turns DCAE-V latents into frames in eight distilled denoising steps.

Deployment is not light. A default setup wants eight GPUs. A single RTX 4090 path exists using an INT8 DiT.

The benchmark and its limits

Hy-Nomas-Bench holds 280 items across four tracks: multi-turn text, camera navigation, interleaved text and camera, and spatial memory. The report compares Hy-Nomas with Genie 3, LingBot-World v2 and Happy Oyster in human pairwise tests. Hy-Nomas wins on interaction and prompt following, with a reported win rate of about 65% on prompt following, and loses to Genie 3 on long-term spatial memory. Those are the authors' comparisons on the authors' benchmark.

Video world models optimise for what a viewer will accept. A driving or robotics stack needs metric accuracy instead: lane widths, lidar returns, a pedestrian who cannot teleport back to the kerb.

The rival it is measured against

Baidu disclosed Apollo World on 3 November. It is an agent-driven simulator that emits camera and lidar observations of a virtual city. In that city, traffic reacts to the vehicle under test. Wang Liang, chief architect of Baidu's Intelligent Driving Group, said a world model should not be a painter producing impressive video but an examiner of driving safety. The "2+3" rubric that followed treats causality and interactivity as gates. It grades physical accuracy, persistence and real-time performance. Baidu published no parameter count, architecture, paper, benchmark table, weights, API or price.

Only one of the two can be downloaded today. Hy-Nomas ships open weights; Apollo World does not.

Neither company has published an independent head-to-head. Until someone outside Tencent runs Hy-Nomas-Bench and publishes the results, the 280 items measure what Tencent chose to measure.