Tencent's Hy-Nomas Steers a 60-Second Video World With Text and Camera on One RTX 4090

Tencent's Hunyuan team released Hy-Nomas on October 21 — an interactive video world model whose community-estimated checkpoint size suggests roughly 8 billion parameters — that generates and extends a scene one short segment at a time while the user keeps steering it with free-form text instructions, camera trajectories, or both. The weights, inference code, training pipeline, and a 280-case benchmark are all published under Apache 2.0, and the model runs on a single GeForce RTX 4090 using INT8 quantization and CPU offloading. A community ComfyUI port appeared the following day. The deployment story is notable, but the more significant detail is buried in the Robotics LoRA that ships alongside the model: Tencent trained it on dual-arm humanoid manipulation data, signaling that Hy-Nomas is being built as a synthetic training environment for physical AI — not merely a video generation tool.

One World, Generated Segment by Segment

The central design choice in Hy-Nomas is what makes it an interactive world model rather than a standard text-to-video system. A conventional generator produces a fixed clip from a single prompt and stops. Hy-Nomas generates a short segment — a few seconds of compressed video latents — then appends that segment's attention keys and values to a permanent, growing KV cache. On the next turn, the user issues a new instruction (change the scene, pan the camera, add an object), and the model generates the following segment while attending to everything it has generated so far. The user steers a world in progress rather than requesting a finished clip.

This approach is called segment-autoregressive generation: autoregressive across segments rather than frame by frame. The benefit is a persistent, steerable world that accumulates history without a separate retrieval system — everything the model has generated stays in its attention context. The cost is that the KV cache and system RAM grow without a hard cap as the session lengthens. Tencent's model card is explicit about this: the CPU offloading path requires system memory to hold the growing history archive, and CPU-to-GPU transfers add latency. Long sessions get expensive in RAM even when GPU VRAM stays bounded by quantization.

Inside the Architecture: Dual Stream, Language Encoder, and Explicit Camera Math

Hy-Nomas pairs a 54-block dual-stream MMDiT backbone with a Qwen2.5-VL-7B-Instruct text encoder — a vision-language model rather than a conventional text encoder, as detailed in Tencent's Hy-Nomas model card. The dual-stream design (descended from Stable Diffusion 3 and Flux) runs separate transformer streams for visual and text tokens through the early layers, joining them for cross-attention. Hidden size is 2,048 with 16 attention heads.

Camera control is not inferred from prose. Hy-Nomas accepts explicit camera-to-world matrices — geometric extrinsics of shape [N, 4, 4] and intrinsics of shape [N, 3, 3] — as a separate conditioning signal. These are normalized, grouped over four video frames, flattened into 64-dimensional vectors per group, and fed into every double-stream block via cross-attention. The separation matters: the text prompt describes what is in the scene; the camera trajectory describes where the virtual lens is pointing. The two compose without competing. A pipeline can supply precise cinematographic paths independently of scene description.

Video is encoded through DCAE-V — 32 channels, 32× spatial compression per spatial dimension, 4× temporal compression — before entering the diffusion process. The distilled model uses 8 denoising steps per segment, enough to generate 720×1280 content at 24 frames per second for playback. Tencent also applies an important attention design the model card calls action-aware causal attention: descriptive context (what the scene looks like) remains accessible across all past segments, but action instructions (the new camera move, the new object) are scoped to their intended temporal windows. Early commands do not accumulate into later segments as persistent noise.

What Tencent Actually Shipped

The release is unusually complete. The Hugging Face repository holds a base checkpoint for community fine-tuning, the distilled production model, an INT8 version for the single-GPU path, a Cinematic LoRA trained on film-quality footage, and a Robotics LoRA trained on the RealSource-World dual-arm manipulation dataset — 35 tasks, 11,428 episodes, roughly 525 gigabytes of real-world robot demonstrations. A fine-tuning pipeline covers supervised fine-tuning data preparation, offline sequence packing for long rollouts, training, and LoRA merging. A pre-built Docker image bundles PyTorch, FlashAttention, and MagiAttention, reducing setup to a single pull. The interactive demo at hy-nomas.hy.tencent.com supports WASD keyboard navigation and mouse-drag camera control. A prompt enhancement service called Nomas-PE rewrites terse instructions into structured generation prompts for the demo; offline inference can load pre-prepared prompt files and skip it entirely.

A community developer of a ComfyUI port measured approximately 21 seconds of generation time per three-second video segment on an RTX 5090. At that rate, generating a 60-second world requires roughly 420 seconds of computation — a workstation-feasible background task rather than a real-time experience on consumer hardware.

Researchers and developers who run the model locally via the Apache 2.0 weights do not route data through Tencent's servers; the online demo at hy-nomas.hy.tencent.com operates separately from offline inference.

How Hy-Nomas Compares to Genie 3

Google DeepMind's Genie 3, released in research preview in August 2025 and made available to Google AI Ultra subscribers in January 2026, operates at the same output specification — 720p at 24 FPS — and is the closest functional peer. The comparison highlights a structural asymmetry: Genie 3's weights are closed, deployment is through Google Labs, fine-tuning is not offered, and local execution is not an option. Hy-Nomas ships weights under Apache 2.0 with a fine-tuning recipe and runs on hardware a developer can buy.

The technical gap, however, runs the other direction on spatial memory. Tencent's pending technical report describes a spatial memory benchmark track — which tests whether generated objects persist correctly after the viewpoint changes and the scene is revisited — and flags this explicitly as an area where Hy-Nomas trails Genie 3. The specific scores from that report remain unverified as of this writing, as the accompanying paper was listed as forthcoming at release. Object permanence across long sessions is a capability Genie 3 developed emergently from training and that DeepMind has specifically highlighted; Hy-Nomas's KV-cache memory approach preserves appearance but does not yet match Genie 3's consistency at viewpoint revisits.

On controllability, Hy-Nomas and Genie 3 occupy different paradigms. Genie 3 supports promptable world events — users interact via text and real-time controls. Hy-Nomas adds explicit camera extrinsics as a first-class input, which suits cinematographic pipelines that supply precise lens paths rather than approximate text descriptions. Matrix-Game 3.0 from Skywork, a 1.8B streaming world model, uses keyboard and mouse signals directly as action inputs — useful for game-controller interaction but not for the text-described or camera-path-described scenes Hy-Nomas targets. NVIDIA's SANA-WM uses six-DoF camera poses with 2.6B parameters across a different hybrid attention architecture; it generates 24 frames per chunk rather than the variable-length segments Hy-Nomas produces.

The Robotics Training Signal

The Robotics LoRA is the detail most coverage has underweighted. The RealSource-World dataset covers dual-arm humanoid robot manipulation — arrangement, grasping, assembly tasks recorded with full sensor annotations — and the LoRA adapter merges directly into the same distilled checkpoint used for world generation, as documented on Tencent's Hy-Nomas model page. This positions Hy-Nomas as more than a visual generation system: a world model fine-tuned on manipulation data can generate plausible arm-and-object sequences, which in turn can serve as synthetic training data for robot policies learning from generated video rather than costly physical demonstrations.

Tencent's Hunyuan team has been developing Hy-Embodied, a line of vision-language-action models for robotics, and has active research hiring specifically for world-model-plus-embodied-intelligence convergence. The Hy-Nomas Robotics LoRA is the data pipeline that connects the world model to that robot stack — a synthetic environment generator that scales training without scaling physical robot time.

What the Release Does Not Resolve

Benchmark scores in Tencent's pending technical report are self-reported by the Tencent Hunyuan team using Hy-Nomas-Bench, a 280-case evaluation dataset that Tencent also published. No independent third-party has reproduced the reported performance figures as of this writing. The benchmark methodology is auditable — the test cases and reference camera trajectories are publicly hosted — but community evaluation will take weeks. A separate benchmark, WorldRoamBench (arxiv:2606.31672), evaluates 12 interactive world models and provides independent context on this generation of systems; its findings on earlier open-source models are worth consulting alongside Tencent's self-reported results.

The missing input modality is image-to-world: every Hy-Nomas session starts from text. Users cannot supply a keyframe, product image, or real photograph as the world's starting point. The distilled model's 196-gigabyte weight set also raises the practical download bar, even for teams with appropriate GPU hardware. And because the Robotics LoRA was trained on data under a CC BY-NC-SA license, the commercial usability of that specific adapter for derived robotic applications warrants legal review.

Where Hy-Nomas goes in the next month will largely be determined by LoRA experiments the community runs on top of the base checkpoint, not by the distilled model's base capabilities. Open weights have a compounding trajectory that closed systems cannot access. Genie 3's spatial memory lead is real and acknowledged; Hy-Nomas's open-weight architecture means that lead is now a target rather than a ceiling.