A world you can interrupt

Tencent Hunyuan released Hy-Nomas earlier this week, making available the weights and inference code for an interactive video world model that generates successive segments under new text instructions and camera trajectories. Most video models take one prompt and return one finished clip. Hy-Nomas splits the job. Hy-Nomas generates video segment by segment, reusing cached attention keys and values from earlier segments as context for subsequent generation. New text instructions and optional camera trajectories condition each segment. In practice that means adding an object, changing the style, describing a new action, or handing it a camera trajectory to follow. The backbone is a 54-block dual-stream diffusion transformer working in a compressed latent space, with a Qwen2.5-VL-7B text encoder reading the instructions and a fine-tuned PixelUMM decoder turning latents back into frames.

That design is aimed at worlds rather than clips, and Tencent ships two adapters that show where it expects the model to be used: a cinematic LoRA for visual quality and a robotics LoRA trained on a real-world robotic-arm dataset. It also ships the test it wants to be judged by. Hy-Nomas-Bench covers multi-turn text interaction, camera navigation, interleaved text and camera control, and spatial memory, in 280 structured cases. What it does not ship is a single result from running that test. Tencent's published materials provide examples and a benchmark dataset, but no benchmark scores or independent evaluations establishing visual quality, spatial consistency or long-session reliability. Demo footage illustrates claimed capabilities; it is not a substitute for reproducible measurements.

What "one RTX 4090" actually buys

The documented configuration deserves credit for being precise enough to argue with. The standard setup is not a consumer setup at all: Tencent's default launch splits the sequence across eight GPUs in one inference group. The single-GPU variant reaches an RTX 4090 by changing three things. The transformer weights and the cached history are stored in INT8, although the main computation stays in BF16. The video decoder runs in FP8. And the text encoder, transformer and decoder take turns on the GPU, returning to the CPU between stages, which the model card notes adds latency and requires enough system RAM for the offloaded models plus a history archive that keeps growing.

The 24 FPS figure describes playback of the generated video, not 24 frames generated per second on an RTX 4090; Tencent provides no measured single-GPU generation throughput.

The sharpest limit is memory of the past. Full-history inference lets each new segment attend to everything generated before it, at a cache cost that grows as the world gets longer. The single-GPU path instead selects up to 18 historical latent frames for GPU attention and archives the rest on the CPU. That budget caps historical context. It does not cap the length of the segment being generated, and it does not cap the memory the current segment needs — peak GPU memory still depends on resolution and segment length. As generation continues, the single-GPU configuration retains a smaller active historical context than the full-history setup. That creates a plausible trade-off for long-range consistency, but Tencent has not published measurements showing how much visual or spatial coherence changes.

The documented single-RTX-4090 configuration covers offline video inference from prepared prompts. The interactive web demo also requires Nomas-PE, a separately served prompt-enhancement component. Tencent additionally provides cinematic and robotics LoRA adapters, but its INT8 inference loader cannot load a separate LoRA adapter at runtime. Those adapters therefore cannot simply be enabled in the documented low-memory configuration.

The licence is the other headline

Set this release next to the open video model readers ask about most. MiniMax H3 is downloadable too, but its Community License bars use and display of the model and its outputs in the United States, the United Kingdom, the European Union and South Korea, and requires separate written authorisation once a commercial product or service built on it generates more than US$20 million in yearly revenue. Tencent licenses Hy-Nomas itself under Apache-2.0, subject to separate licenses for third-party components. Unlike the MiniMax H3 Community License terms previously examined by ALFW, Hy-Nomas's stated model license does not impose a geographic output restriction or revenue-based commercial authorization threshold. That makes its stated licensing terms less restrictive for commercial evaluation and deployment, although teams still need to review third-party component licenses.

What Apache-2.0 cannot settle is whether the model is any good. On that question the release is unusually quiet: a benchmark with no scores, a report that has not arrived, and a frame rate that describes the output file rather than the wait. The download is real, the one-card configuration is documented within its stated bounds, and the licence means what it says. The evidence that it produces worlds worth keeping is still outstanding.

Sources

  1. [1] Tencent Hunyuan, Hy-Nomas model card — Hugging Face (read first-hand by ALFW) — architecture, the standard eight-GPU and single-RTX-4090 configurations, runtime considerations and the Apache-2.0 licenceRead source
  2. [2] MiniMax, MiniMax H3 Community License Agreement — Hugging Face licence file (re-read first-hand by ALFW for this article) — the Excluded Territories (the European Union, the United Kingdom, South Korea and the United States) and the US$20 million yearly-revenue authorisation thresholdRead source
  3. [3] Superpower Daily, “Tencent releases interactive video AI with a single-RTX-4090 configuration” (October 2026) — summary of the release and its single-GPU caveats, sourced to the model cardRead source
  4. [4] FourWeekMBA, “Moonshot Prices Kimi K3.1 Pro at Half of Claude Opus 5.5” (October 2026) — model page created 21 October and updated 23 October; no benchmark results on the cardRead source
  5. [5] OrcaRouter, “Hi3D v1.0 vs Hy-Nomas: Apache 2.0 vs a 200k-MAU Licence” (October 2026) — licence and hardware comparison; neither publisher has released evidence supporting a quality comparisonRead source