Hy-Nomas vs Hunyuan3D WorldClaw: Same Lab, Two Kinds of World

Tencent's Hunyuan group shipped two things called a world model within three months of each other, and they disagree about almost everything. Hunyuan3D WorldClaw, published in August 2026, is not really a model at all: it is a four-stage agentic pipeline that plans a scene with LLM agents, generates terrain, reconstructs individual objects as textured meshes, and then has a vision agent render the result and fix what looks wrong. Hy-Nomas, released on 23 October 2026, is a single segment-autoregressive diffusion transformer that renders steerable video and ships its weights under Apache 2.0. One builds an editable 3D asset from a sentence. The other renders a continuous world from a conversation.

The comparison is worth making precisely because they share a lab. It is a clean natural experiment in two strategies: orchestrate frontier models into a pipeline, or train one model to do the whole thing. The two releases also invert each other on openness — the one with the heavier engineering stack published no code, and the one that is a single model published everything.

A model and a pipeline are different kinds of artifact

WorldClaw's paper describes a router. The system orchestrates several frontier models rather than performing generation itself: LLM agents for planning, an image model for regional composites, a segmentation model, Tencent's own Hunyuan3D for mesh reconstruction, and Blender driven through MCP for assembly. The paper names Claude Opus 4.8 as the planner and GPT-Image-2 and SAM3D among the regional components. Its contribution is the choreography — the global-to-regional decomposition and the render-based repair loop — not a new generative backbone.

Hy-Nomas is the opposite bet. One 54-block dual-stream MMDiT, hidden size 2,048, a Qwen2.5-VL-7B-Instruct text encoder, camera conditioning through cross-attention, and a lifelong KV cache. There is no planner, no segmentation stage, no external renderer in the loop. The model is the system.

That difference explains the failure modes. A pipeline inherits the weaknesses of every component it calls and can be repaired one component at a time. A single model has one coherent world view and no seam between stages, but any weakness is baked into the weights and can only be fixed by retraining or by a LoRA.

What WorldClaw's four stages actually do

Stage one is intent analysis and scene planning. Two LLM agents split the job: one extracts only the constraints the user actually stated, the other fills in the details to produce a structured scene specification covering regions, terrain, objects and spatial relations. The split is a sensible defence against the usual failure of a single planner, which is inventing requirements and then satisfying them.

Stage two generates global terrain: a semantic layout map, region-aware height fields, and materials, rocks and vegetation assigned per region. Stage three renders each region into a 2D composite, segments the instances, reconstructs each one as a textured mesh through Hunyuan3D, and places it back on the terrain. Stage four is the interesting one — a vision agent renders the assembled scene from multiple viewpoints and looks for floating objects, interpenetration, wrong scale and wrong pose, then iterates. Assembly happens in Blender through MCP.

The output is explicit, textured, individually editable meshes that export to a game engine. That is a genuinely different deliverable from a video: every tree is an object, every object has a transform, and a level designer can move one without re-rendering the world.

Meshes you can edit, or video you can steer

Hy-Nomas's output is 720 × 1280 video at 24 FPS, generated in segments with eight denoising steps in the distilled configuration. You cannot open it in Blender. You can, however, redirect it mid-session: free-form text for what is in the world, camera extrinsics for where the shot goes, and a KV cache that keeps the earlier segments consistent with the new instruction. The edit is conversational and happens before the next segment renders.

WorldClaw's output is editable after the fact. You get geometry, and the edit is spatial — move the object, rescale the tree, re-render the region. What you cannot do is hold a conversation with the world while it renders. The pipeline runs to completion, then you work on the asset.

• Output — Hy-Nomas steerable 720 × 1280 video vs Hunyuan3D WorldClaw explicit textured meshes in a scene

• Unit of work — Hy-Nomas a segment rendered on request vs Hunyuan3D WorldClaw a whole world generated in four stages

• Editing — Hy-Nomas text and camera instructions between segments vs Hunyuan3D WorldClaw direct manipulation of meshes after generation

• Engine — Hy-Nomas one diffusion transformer vs Hunyuan3D WorldClaw an orchestrated stack of frontier models plus Blender

• Hardware — Hy-Nomas 8 GPUs by default, one RTX 4090 in the INT8 low-memory path vs Hunyuan3D WorldClaw 4 × NVIDIA H20

• Release — Hy-Nomas Apache 2.0 weights, code and benchmark dataset vs Hunyuan3D WorldClaw a paper, a project page and a README with no code or weights

The cost of an agentic world

WorldClaw's compute story is not a GPU-hour number in the usual sense. It is four H20s plus a large number of frontier-model calls: planning tokens, image generation per region, segmentation per instance, mesh reconstruction per object, and repeated vision-agent renders during refinement. The marginal cost of a world is dominated by those calls, and the paper does not publish a price. Community rebuilds that substitute hosted image and 3D endpoints report roughly fifty cents and a few minutes per world, which is a useful order of magnitude but not a vendor figure.

Hy-Nomas's cost is more conventional: eight GPUs with eight-way sequence parallelism by default, or one RTX 4090 with INT8 DiT weights, an INT8 KV cache, sliding-window history, CPU offload, FlexAttention and a CUTLASS FP8 decoder. 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. No hosted price exists because there is no hosted service.

The comparison that matters is not dollars per world but where the money goes. WorldClaw spends on tokens and API calls, which fall as frontier-model prices fall. Hy-Nomas spends on GPU occupancy, which does not. A pipeline gets cheaper on someone else's curve; a model gets cheaper only when someone distils it.

• Hy-Nomas cost driver — GPU hours per generation session, no per-call billing

• WorldClaw cost driver — frontier-model calls for planning, imaging, segmentation, reconstruction and repair, plus 4 × H20

• Hy-Nomas scaling lever — distillation and LoRA, both already in the release

• WorldClaw scaling lever — cheaper component models and fewer repair iterations

WorldClaw is also the clearest argument for a routing layer that this batch contains, because it is one. Its planner calls one model, its regional stage calls another, its reconstruction stage calls a third, and the whole thing is a routing policy with a renderer attached. OrcaRouter does the same job for hosted models — 200+ models behind one key, provider list price passed through at 0% markup, automatic failover, and a routing DSL that composes several models into one call, with model fusion available when you want a panel rather than a chain. Neither Hy-Nomas nor WorldClaw is on our catalogue, and we are not going to pretend a self-hosted world model is an endpoint; what we route is the layer WorldClaw demonstrates you need.

Openness is inverted

This is the strangest part of the pair. WorldClaw is the more elaborate system — a multi-stage agentic pipeline with a published paper and a project page — and the official GitHub repository contains a README, the paper and project assets. No code, no weights, no API, no pricing. An issue asking for the code has gone unanswered. What exists publicly are community reimplementations that rebuild the pipeline on top of hosted endpoints, which is a testament to the paper's clarity and not a substitute for the release.

Hy-Nomas is the simpler system and the more open one: Apache 2.0, full checkpoints on Hugging Face, inference and training code on GitHub, a captioning harness, a LoRA trainer, a merge script, and the 280-case benchmark dataset. The technical report is missing and the benchmark has no scores, but the artifact is complete enough to reproduce and modify.

For anyone planning to build, that inversion decides more than the architecture does. A pipeline you cannot run is a design document. A model you can run is a starting point, even when its evaluation is a blank page.

Which to reach for

Reach for Hunyuan3D WorldClaw's approach when the deliverable is a level: a large, explorable, editable 3D environment that has to go into a game engine and be modified by a human. Be aware that you will be rebuilding it from the paper, because Tencent has not released it, and that its quality claims are qualitative — the paper compares against SynCity, World Labs' Marble, MajutsuCity, Meta's WorldGen and GPT-5.6 Sol without FID, CLIP or a user study.

Reach for Hy-Nomas when the deliverable is footage or a training environment: pre-vis, synthetic video, a robotics scene generator you intend to adapt with the shipped Robotics LoRA. Be aware that the long-horizon consistency claim is a design argument plus a demo, not a score, and that a single 4090 run is possible but carries offload overhead.

And if you need both — an editable world and steerable footage of it — the honest answer is that no single Tencent release gives you that today. WorldClaw produces geometry you cannot animate conversationally. Hy-Nomas produces video you cannot export. The gap between them is the product nobody has shipped yet, and it is a more interesting thing to watch than either model on its own.