Hy-Nomas vs Odyssey-3: Open Weights Without Scores, Scores Without Weights

Ask Odyssey for Odyssey-3 and you get a number: 66.1 on Physics-IQ Verified, the highest raw video-to-video score the company reports. Ask for the checkpoint and you get a contact form. Ask Tencent Hunyuan for Hy-Nomas and the exchange runs the other way: roughly 196 GB of Apache 2.0 weights, an inference harness, a LoRA trainer and a browser demo, with no score attached to any of it. The two models occupy the same category, interactive autoregressive diffusion-transformer world models, and the same month. Odyssey-3's public research preview opened on 8 October 2026; Hy-Nomas's weights went up between 21 and 23 October 2026. But the decision between them is not a question of which one is better, because neither side has published the evidence that would settle it, and each is missing a different half.

What follows is only what can be established from the two releases: how each model treats the camera, what the numbers on each side do and do not measure, and what it costs to get either one running in front of a user.

The box, not the architecture

Both are autoregressive diffusion transformers, and at the level of block diagrams they are cousins. Hy-Nomas renders a video segment, appends it to the session, and conditions the next segment on a lifelong KV cache; the backbone is a 54-block dual-stream MMDiT at hidden size 2,048 with 16 attention heads and 128 dimensions per head, a Qwen2.5-VL-7B-Instruct text encoder and 32-channel DCAE-V latents. Odyssey-3 is described by its maker as a multi-step video diffusion transformer extended autoregressively with teacher forcing and causal masking, then distilled for real-time interaction. Both decide to render the world in pieces and to remember what came before.

The difference that matters is what is inside the release. Hy-Nomas ships a kit: base DiT in FP32, a distilled DiT, an INT8 build, the DCAE-V autoencoder, a PixelUMM decoder in FP32 and FP8, the Nomas-PE prompt enhancer, two LoRA adapters, a Docker image, a captioning harness for building fine-tuning data, offline sequence packing, a LoRA trainer and a merge script. Odyssey-3 ships a browser tab and a queue.

• Weights — Hy-Nomas Apache 2.0, about 196 GB; Odyssey-3 none released and no license published

• Inference code — Hy-Nomas repository plus a Docker image; Odyssey-3 none

• Fine-tuning — Hy-Nomas ships the LoRA trainer and a documented recipe; Odyssey-3 not offered

• Parameter count — neither vendor discloses one; the Hy-Nomas FP32 checkpoint is 34.2 GB, which implies roughly 8.5 billion parameters if you assume 32-bit storage and no optimizer state, an arithmetic estimate rather than a specification

• Access — Hy-Nomas self-host or Tencent's demo; Odyssey-3 browser preview, API by contact form

Who decides where the camera goes

Odyssey-3's interaction model is navigation. The preview supports first-person movement, third-person movement and independent camera movement, and the world keeps predicting while you do it. That is the right design for a person exploring an environment, and for the demo reel that sold the model: you walk, the world reacts, the illusion holds.

Hy-Nomas's interaction model is a trajectory. You supply camera extrinsics, a position and orientation path, and the model renders that path. Text and camera are separate channels: the prompt says what is in the scene, the trajectory says where the lens is. The two compose rather than compete. Nomas-PE exists to rewrite a terse instruction into something the model conditions on more effectively, which tells you something about the text channel's sharp edges.

The practical consequence is reproducibility. A navigation session is a traversal: real-time, person-driven, and awkward to replay exactly. A trajectory is a parameter you can log, vary, and regenerate against. If you are producing synthetic training data for a robot or an agent, the second is an annotation and the first is a demonstration. That is not a claim that Hy-Nomas looks better. There is no evidence it does. It is a claim about which control surface a pipeline can consume.

The resolution line, and why not to lead with it

The obvious line is that Hy-Nomas renders 720 × 1280 at 24 FPS while Odyssey-3 renders 832 × 480, with a Pro tier at 1280 × 720. That comparison is true and nearly meaningless on its own. The 24 FPS figure is playback of a generated segment, not generation speed. Odyssey-3's 16 FPS is measured on the Physics-IQ harness, and the frame rate of the Pro tier is not published at all. Neither vendor has shown what its output looks like next to the other's at the same resolution, because they have never been run on the same task.

• Base resolution — Hy-Nomas 720 × 1280, overridable to 480 × 832; Odyssey-3 832 × 480

• Top tier — Hy-Nomas has one configuration; Odyssey-3 Pro runs at 1280 × 720

• Frame rate — Hy-Nomas 24 FPS playback with generation speed unstated; Odyssey-3 16 FPS on the benchmark harness, Pro unstated

• Denoising — Hy-Nomas 8 steps in the distilled configuration, with 4 and 6 also accepted; Odyssey-3 is distilled for real time but its step count is not published

One leaderboard, two ways of not answering it

Physics-IQ Verified, run by Anates Labs with Google DeepMind, is the closest thing either model has to independent scrutiny. It asks a model to continue video of a real physics experiment and scores the continuation against what actually happened. Forty-one models from sixteen labs are on the board. Odyssey-3 Pro's 66.1 is the figure the company leads with, and it is the highest raw video-to-video score reported.

Read the board itself and the picture gets more specific. It ranks by net improvement over the track mean, not by raw score, and on that column FLUX 3 [large] sits first at +12.27 points, Odyssey-3 Pro second at +11.15, and the base Odyssey-3 third at +9.99. Odyssey's submissions appear twice under two prompt regimes: with custom prompts, +11.15 and +9.99; with the benchmark's own prompts, the base model falls to +2.15 and eighth place. The 66.1 comes from best-of-8 sampling; a single generation scores 63.37. The board is a submission board, not a third-party run.

None of that makes Odyssey-3 a bad model. It makes the headline a selected configuration, and the leaderboard's own ranking a different claim from the one in the launch copy. A reader who quotes 66.1 and a reader who quotes rank 02 are both quoting the same board, and only one of them is quoting how the board scores.

Hy-Nomas's side of the evidence is simpler and emptier. Tencent published Hy-Nomas-Bench: 280 cases across four tracks, covering multi-turn text control, camera navigation, interleaved text-and-camera control and spatial memory, with about 220 reference camera trajectories. The dataset is real and downloadable. The results table does not exist. The technical report link on the model card says "Coming Soon" and resolves to nothing.

So the honest state of this comparison, sixteen days after Odyssey-3's preview opened and three days after Hy-Nomas's weights appeared, is that the two models have never been evaluated on the same task, by the same harness, at the same settings. Physics-IQ does not list Hy-Nomas. Hy-Nomas-Bench does not list Odyssey-3. Anyone who tells you which is better is extrapolating from architecture and demo clips.

What it costs to touch

Odyssey-3 has no published price. The Physics-IQ board carries a normalized cost estimate of $0.139 per generated video for the base model and $0.267 for Pro, normalized where possible to 24 fps and 1280-wide output, but those are benchmark accounting figures, not rates you can buy at. Access is the browser preview and a contact form for API priority. The site's API terms still describe an Odyssey-2 prototype, which is a fair signal of how far the commercial layer has moved.

Hy-Nomas has no price either, because there is nothing to buy. Tencent does not operate a hosted endpoint for it. The cost is hardware: eight GPUs with eight-way sequence parallelism on the default path, or a single RTX 4090 using INT8 DiT weights, an INT8 KV cache, an FP8 decoder, a sliding-window history cap and CPU offload. The full checkpoint set is about 196 GB, the documented demo path pulls about 86 GB, and the low-memory path needs roughly 38 GB plus the text encoder. The training example in the repository uses 128 GPUs as sixteen data-parallel replicas across eight sequence-parallel ranks.

Put those side by side and the trade is not close on convenience: Odyssey-3 costs a browser tab today, Hy-Nomas costs a GPU and a weekend. The weekend buys something a browser tab never will, which is a checkpoint you can fine-tune on your own footage, merge into a distilled build, and serve without asking anyone's permission.

Where the routing layer actually belongs

Neither model is on our catalogue. Hy-Nomas is open weights you host yourself; Odyssey-3 is a closed preview behind a contact form. We are not going to imply otherwise, and there is no model page of ours to link for either.

The reason routing still comes up is the loop around the model rather than the model itself. A Hy-Nomas session needs something to decide the next instruction and, often, to rewrite the prompt before Nomas-PE sees it. An Odyssey-3 prototype needs something to log the session, judge whether the environment stayed coherent, and turn a traversal into a structured evaluation. Those are ordinary language-model calls, and they are the ones that fail at three in the morning. OrcaRouter puts 200+ hosted models behind one key at 0% markup, passes provider list price through unchanged, and lets you compose the prompt-rewriter, the judge and the fallback into one routing DSL chain instead of three vendor contracts. The world model stays where it is; the scaffolding around it gets one endpoint and automatic failover.

Three decisions, three answers

• You need a model you can fine-tune on proprietary footage, ship on your own hardware, or merge a domain LoRA into — Hy-Nomas, with no serious alternative in this pair. The documented recipe, rank and alpha 256 under Muon at a constant 1e-4 for 2,000 steps, trained from the FP32 checkpoint and merged into the distilled one, is specific enough to run this week.

• You need an explorable environment in front of a stakeholder this afternoon and you do not need to own it — Odyssey-3's preview. It is also the only one of the two with a submission on a third-party board, caveats included.

• You need a defensible answer to "which is better" for a procurement document — neither. That document does not have a source yet.

What would change the answer

Three things, in order of likelihood. A Hy-Nomas technical report with actual numbers would end the one-sided evidence problem and turn the next version of this page into a real comparison. A community LoRA wave would show whether the fine-tuning recipe works outside Tencent's own examples, which is the claim with the shortest path to being tested. An Odyssey-3 weights release, or even a documented API price, would move it from a demo you visit to a component you can plan around.

Until one of those lands, the useful way to read this pair is by what each one lets you keep. Odyssey-3 lets you keep an impression. Hy-Nomas lets you keep the model. Which of those you need is a question about your project, not about the benchmark.