Hy-Nomas: Tencent Open-Sourced an Interactive World Model and Said Nothing About It

Tencent put the evaluation suite and the weights for Hy-Nomas live this morning and, by the end of the day, still had not published a single score from either. That is the shape of this release. Hy-Nomas is a video world model you steer with text instructions and camera moves, one segment at a time, and its weights went up on Hugging Face at 11:18 UTC today, October 23, 2026: one commit carrying 196 GB of Apache-2.0 checkpoints, alongside a code repository pushed at 09:01 the same morning and a playable demo. Two days earlier both repositories existed and were empty. There is no blog post, no arXiv entry, and the "Technical Report (Coming Soon)" badge on the model card links to a file that returns 404. Hours after upload the model page shows zero likes and zero downloads, and the only sign anyone has noticed is the GitHub repository: eight stars, one commit, no description.

So the honest framing is not "Tencent launched a world model." It is: Tencent published the weights, the training and inference code, a benchmark, and a demo site, and has not yet said a word about what any of it scores. Everything below is either a timestamp or file size you can verify yourself, a number Tencent typed onto its own card, or a place where the card is conspicuously silent. The three categories are kept apart deliberately, because in a release like this the silence carries as much information as the specs.

What actually landed, to the hour

The artifacts are easy to date because there are so few of them and none has been edited much.

• The model repository tencent/Hy-Nomas was created on Hugging Face on 2026-10-21 at 02:17:37 UTC and sat empty until its only commit landed on 2026-10-23 at 11:18:04 UTC. That commit is 196.4 GB across 46 files: two 34.2 GB fp32 DiT checkpoints, a 9.06 GB INT8 DiT, two LoRA adapters, a 31.55 GB PixelUMM decoder, its 9.13 GB FP8 twin, a 4.28 GB DCAE-V autoencoder, an 18-shard prompt-enhancement model totalling 55.6 GB, and the usual tokenizer and config furniture.

• The code repository Tencent-Hunyuan/Hy-Nomas was created 47 minutes later, at 03:04:54 UTC, with no description field filled in. Its single commit was authored on October 21 but only pushed today, at 09:01:42 UTC. The repository still shows one commit and no releases, and it holds the training code, inference scripts, a Docker image and a web-demo deployment guide.

• The benchmark dataset tencent/Hy-Nomas-Bench was created on 2026-10-22 at 05:18:20 UTC and received its one and only commit today at 09:10:51 UTC.

• The demo at hy-nomas.hy.tencent.com answers with HTTP 200 and a page title. It is the one part of this release aimed at anyone outside a research group.

The quietness is specific to this model, not to Tencent. The company's own Hugging Face organisation is busy: tencent/Hy4-preview from late August has 513 likes and roughly 18,400 downloads, and tencent/Hy-Memory-30B-A3B from October 15 has 297 likes and about 3,300 downloads inside eight days. Hy-Nomas, uploaded in the same window and vastly larger, has none of that. Whatever the internal reasons, this is not a launch that underperformed. It is a launch that was not attempted.

What Hy-Nomas does that a text-to-video model does not

The distinction matters more than the parameter counts, because it is the reason to care about the release at all. An ordinary text-to-video model takes one prompt and returns one clip; if you want something different you start again and get an unrelated clip. Hy-Nomas generates a world as a chain of segments and keeps its own history.

Concretely, the card describes four things happening at once. Each segment gets its own prompt, so you can introduce an object, change the style, alter the setting or describe an action partway through generation rather than up front. A persistent key/value cache carries the generated history forward, so a later segment is conditioned on what was already produced instead of being invented from scratch. Camera motion is a separate conditioning path rather than something you describe in prose: you supply camera-to-world matrices and intrinsics, and the model attends to them in every block. And because segments are generated autoregressively, the world can keep extending for as long as you keep feeding it instructions.

The prompt format shows how literal this is. Offline inference takes a JSON list where each entry has a style, a perspective, a subject list, and then timed segments — a start and end time, a long visual description, and contained actions with their own timestamps. The first segment establishes the world; later segments use a description_change field to say how it should evolve. The bundled example builds a luminous forest around a small orange fox, then opens glowing blue flowers beside its path at the two-second mark.

One number in the card is easy to misread, and the card itself flags it: the 24 FPS figure is the playback rate of the output video, not generation speed. How fast the model actually runs depends on hardware, resolution, segment length and how much history is in the cache. Tencent says as much directly. In a release with no published latency and no published scores, that is the only honest reading of the number.

The spec sheet the card does give

Tencent is unusually precise about architecture, which is what makes the missing results feel pointed.

• Backbone — 54 dual-stream MMDiT blocks and zero single-stream blocks, per the inference configuration.

• Width — hidden size 2,048, with 16 attention heads at 128 dimensions each and a 4× feed-forward expansion to 8,192.

• Position encoding — 3D RoPE with dimensions [16, 56, 56] and theta 256.

• Text path — the language-model component of Qwen2.5-VL-7B-Instruct, whose 3,584-dimensional hidden states pass through a two-layer token refiner and project into the 2,048-dimensional DiT space.

• Camera path — normalised 4×4 extrinsics grouped across four frames and flattened to 64 dimensions, applied by cross-attention in every block.

• Latents — DCAE-V f32t4c32: 32 channels, 32× spatial compression, 4× temporal compression.

• Decoder — a fine-tuned PixelUMM with 36 layers, 4,096 hidden size, 32 query heads and 8 KV heads, running one decoding step over 10-latent-frame windows with 5-frame overlap.

• Output — 720 × 1280 at 24 FPS, with 8 denoising steps in the distilled configuration; the code repository says the distilled checkpoints support 4, 6 or 8 steps.

• Licence — Apache 2.0, with the Qwen text encoder staying under its own terms.

The one figure Tencent never states is total parameter count. The number that comes closest is derived rather than published: a 34.2 GB fp32 DiT checkpoint works out to roughly 8.5 billion parameters at four bytes each, and the 9.06 GB INT8 build is consistent with that. Treat it as arithmetic on a file size, not as a spec sheet claim — the vendor has not confirmed it anywhere.

What it costs to run, which is the real gate

Nothing about this release is hostile to outsiders, but the default deployment assumes you have a node, not a laptop. The standard inference configuration uses 8-way sequence parallelism, so its default launch wants eight GPUs in one inference group. The README is upfront about that rather than burying it.

There is a genuine single-GPU path, and it is the most interesting engineering in the release. It runs on one RTX 4090 by combining four tricks: DiT weights and cached keys and values are stored in INT8 while computation stays in BF16, the decoder runs in FP8 via CUTLASS, the text encoder and both networks shuffle between CPU and GPU for their own stages, and active GPU history is capped at 18 latent frames with the older history archived in system RAM. That last trade is the one to understand before planning around it: the archive keeps growing, the 18-frame budget only bounds what sits on the GPU, and the current segment's memory is not capped by it at all. Tencent says plainly that CPU–GPU transfers add latency and that peak GPU memory varies with resolution and segment length.

The download maths follows from the file sizes. The documented demo path needs the distilled DiT (34.2 GB), the matching PixelUMM decoder (31.55 GB), DCAE-V (4.28 GB) and roughly 16 GB more for the Qwen encoder — about 86 GB before you generate anything. The single-GPU path swaps in the INT8 DiT and FP8 decoder for a far smaller footprint, near 38 GB plus the encoder. The two LoRA adapters are separate purchases of disk: a 6.12 GB cinematic adapter and a 12.24 GB robotics adapter trained on the RealSource-World dual-arm dataset, both meant to be merged into a floating-point DiT. The INT8 loader cannot take a separate LoRA at inference time, so those two paths do not compose.

Environment requirements are equally specific: Linux, Python 3.12 or newer, CUDA 13.0 Toolkit, PyTorch 2.10.0, and builds of FlashAttention 2, FlashAttention 3 and MagiAttention. A prebuilt image, hynomasoss/hynomasoss:cu130, is recommended over the manual path — with attention-kernel compilation as the reason. The training configuration in the code repository goes further: DP16 × SP8, 128 GPUs, sequences packed to 360,000 tokens. Whatever you do with the weights, you are not retraining this at home.

The benchmark Tencent shipped without scores

Hy-Nomas-Bench is not an afterthought. It is a real dataset: four evaluation suites covering multi-turn free-form text interaction, camera navigation, interleaved text and camera control, and spatial memory, with 280 test cases, 220 reference camera trajectories, structured prompts, interaction sequences and keyboard timelines, released as JSON under Apache 2.0 and organised into evaluation-only splits. It landed roughly two hours before the weights did.

What it does not come with is a results table. There is no leaderboard, no per-suite score, no comparison against any other model, and no indication that one is coming. And the card is careful in a way that reads as deliberate: after listing architecture and sampling defaults, it adds one line — "Architecture and sampling defaults are configuration details, rather than benchmark results."

That sentence is doing real work. It means the eight denoising steps, the 720 × 1280 resolution and the camera guidance scale of 1.0 are not evidence of quality, and Tencent is declining to let them be read as such. For anyone evaluating the model, the practical consequence is simple: you can run the evaluation yourself, and until then you cannot put Hy-Nomas next to anything else on a number. The suite is the thing to download if you want a verdict, because nobody is handing you one.

Where it sits among the other open world models

The field moved fast in 2026, and the comparison that matters is not parameter count but what evidence each release brought with it.

• Weights and licence — Hy-Nomas is Apache 2.0 with no gating. LingBot-World, from Ant Group, shipped open weights in late January 2026 and was pitched as the first open model comparable to Google's closed Genie 3.

• Control scheme — Hy-Nomas is text-first with camera trajectories, segment by segment. LingBot-World and the Genie 3 line lead with real-time keyboard and mouse control of a character or camera, which is a different interaction model: continuous play versus directed construction.

• Hardware floor — Hy-Nomas wants eight GPUs by default or one RTX 4090 with a capped history. Astronex-World 1.0, published in September 2026, reported streaming in real time on a single 48 GB L20 at 832 × 480, which is a markedly lower bar.

• Evidence — this is where Hy-Nomas stands alone in the wrong column. LingBot-World shipped with reported throughput and latency figures; Astronex-World shipped with an arXiv paper and benchmark numbers. Hy-Nomas has a 404 on its technical report and no scores at all.

• Lineage — Tencent's previous open interactive effort, HY-WorldPlay, worked from a starting image plus pose and action input at 720p and 24 FPS. Hy-Nomas is not the same model with a new name; the text-multi-turn loop and the per-segment prompt structure are new, and the card does not claim descent from it.

None of which settles quality. A model can be excellent and undocumented, and Tencent's silence is not evidence either way — it just means the burden of proof has moved to whoever downloads it.

Who should touch it now, and who should wait

If you have an eight-GPU node and your problem is world state that a human edits mid-generation — a scene that gains an object at second four, a camera that turns because you told it to, a world that keeps its geometry while you change the weather — this is the most direct open implementation of that idea currently downloadable, and the Apache 2.0 terms mean you can build on it without a negotiation. The robotics LoRA is a signal of where Tencent expects it to be used: simulation worlds for embodied training, not social video.

If you need a number before you commit, wait. Not because the model is suspect, but because there is nothing to cite. A benchmark you have to run yourself is a week of engineering, and the technical report, when it appears, will probably change what you would have measured.

There is a smaller operational question around the model that is worth naming, because it bites anyone who tries to reproduce the demo. The web demo routes every instruction through Nomas-PE, a separate 55.6 GB service that converts loose user phrasing into the structured segment JSON the DiT expects, and Tencent states that the demo requires it. Offline inference skips that step and asks you to write the structured prompts yourself. If you are building your own front end, that prompt-construction layer is an ordinary language-model call — and that part, unlike the video stack, is commodity infrastructure. OrcaRouter does not serve Hy-Nomas — a 196 GB self-hosted diffusion stack is not what an API route is for. What it routes is the text side around it — one key for 200-plus models at 0% markup on provider list price, with automatic failover — so a prompt-rewriting or output-captioning step in front of a self-hosted video model does not need its own vendor contract.

The open questions, and what would change this article

Four things are genuinely unknown, and each has a specific artifact that would settle it.

The first is quality. Hy-Nomas-Bench exists, so a published results table is the obvious next step, and its absence is the single largest gap between this release and a citable one.

The second is speed. No latency figure appears anywhere in the card, only the playback rate. Someone will have to measure time-to-first-segment on real hardware, and the eight-step distilled configuration versus the base autoregressive model is a difference worth quantifying rather than assuming.

The third is ecosystem support. The repository carries a diffusers tag, but the documented path is Tencent's own hymm package and a purpose-built Docker image; no third-party runtime is named as supported. Whether this gets a pipeline class in a mainstream library, the way most open video models eventually do, decides whether it stays a research artifact or becomes a tool.

The fourth is whether Tencent ever says anything. The technical report is labelled coming soon and currently 404s, there is no arXiv preprint, and no press coverage exists as of October 23. It is entirely possible the company is waiting on a paper or a hosted product before it promotes any of this. It is also possible the release is exactly what it looks like: weights published for the record, with the announcement deferred.

Either way, the download is live, the licence is permissive, and the demo takes instructions today. For a model nobody announced, that is more than most quiet releases offer — and considerably less than a launch.