The editorial argues Hy3's key differentiator is clean topology output that can drop into Blender or Unity without a full retopology pass. Prior open efforts like Shap-E, Point-E, and Gaussian Splatting variants looked fine in viewers but broke when rigged, UV-unwrapped, or baked — Hy3 is the first open release to clear that bar.
Multiple commenters concede Meshy and Luma still win on cherry-picked best-case demos, but argue Hy3 produces better median-quality output. That distinction matters when you're generating thousands of assets in a pipeline and can't hand-pick winners.
The HN thread splits between users who ran Hy3 locally and were impressed by the geometry quality and those who hit VRAM walls before getting output. Both camps agree it's a step up from Hunyuan3D-2, but accessibility depends heavily on available GPU memory.
The editorial frames 3D as generative AI's awkward middle child — image models crossed the production line at SDXL, video at Sora/Kling, but 3D has been stuck with demo-quality outputs requiring hours of cleanup. Hy3 represents the moment the open 3D stack catches up, breaking the closed vendors' per-asset pricing model that only worked for hero pieces.
Tencent's Hunyuan research team published Hy3, a text-and-image-to-3D generation model, at hy.tencent.com/research/hy3. The post hit 471 points on Hacker News within a day, which is where most of the developer discussion is happening. The release includes model weights, an inference pipeline, and a gallery of generated meshes — the usual open-weight package, but from a Chinese hyperscaler that has historically been less generous about license terms than Meta or Mistral.
The headline capability is a single-pass generation of textured 3D meshes from a text prompt or a single reference image, with topology clean enough to drop into Blender or Unity without a full retopology pass. That last part is what separates this from the last two years of "here's a 3D model, good luck getting it into your engine" releases. Earlier open efforts — Shap-E, Point-E, various Gaussian Splatting variants — produced outputs that looked fine in a viewer and fell apart the moment you tried to rig, UV-unwrap, or bake lighting.
The HN thread is split between people who ran it locally and came back impressed, and people who ran it locally and hit VRAM walls. Both groups agree the output geometry is a clear step up from Hunyuan3D-2, Tencent's previous drop from earlier this cycle. Nobody in the thread is claiming it beats closed commercial offerings from Meshy or Luma on the best-case demos — but multiple commenters point out it beats them on the median case, which is what actually matters when you're generating a thousand assets and can't cherry-pick.
Three-D generation has been the awkward middle child of generative AI. Image models crossed the "usable in production" line around SDXL. Video models crossed it somewhere around Sora and Kling. Three-D has been stuck in a valley where the demos look great, the outputs need six hours of cleanup, and the closed vendors charge per-asset pricing that only works if you're generating a handful of hero pieces.
What Hy3 changes is the economics of bulk 3D asset generation. If the weights actually run on a single H100 — the technical report claims they do, and community reproductions are backing it up — then a small studio can generate a thousand assets overnight for the electricity cost, not the API cost. That's the threshold that turns 3D generation from a marketing gimmick into an actual pipeline component.
The comparison worth drawing isn't to other 3D models. It's to what happened when Stable Diffusion 1.5 shipped. Closed vendors had better outputs on carefully-curated prompts. Open weights won because they were good enough on the long tail, and the long tail is where real products live. Meshy and Luma are Midjourney in this analogy — better polish, faster iteration on their best cases, but expensive per-unit and locked into their pipelines. Hy3 is the open baseline that resets everyone else's pricing conversation.
The skeptical read, which some HN commenters are making, is that Tencent has released open weights before and quietly walked back the licenses when the models got traction. The Hunyuan-Video license had commercial-use restrictions that surprised people who assumed "open" meant Apache. Read the Hy3 license before you build a product on it — the previous Hunyuan releases carried region and revenue-tier restrictions that would break most Western SaaS assumptions. The community reaction so far suggests the terms are more permissive this time, but "more permissive than Hunyuan-Video" is a low bar.
Community reactions from the HN thread cluster around three points. Game dev commenters are excited about the topology quality and asking whether it handles hard-surface modeling (early answer: better than expected, still weaker than organic shapes). VFX and product-viz people are asking about UV quality and PBR material output (mixed — UVs are usable, materials need work). And infrastructure-minded commenters are noting that the inference stack is straightforward PyTorch, no exotic CUDA kernels, which means it'll port to consumer GPUs within a week.
If you're currently paying Meshy or Luma per-asset for programmatic 3D generation, run the Hy3 evaluation this week. The bar isn't "is it as good as the vendor." The bar is "is it good enough that self-hosting saves you money at your current volume." For most workloads, that answer flipped on Tuesday.
If you're building any product that involves user-generated 3D content — avatars, storefronts, game mods, AR try-on — Hy3 is now the baseline your closed vendor has to beat, and you should be renegotiating. The leverage in vendor conversations after a credible open release is significant, and it lasts about six weeks before pricing hardens around the new floor.
If you're in a compliance-sensitive environment, hold off. The license terms and the provenance of training data are worth a legal review before you ship anything Hy3-derived to end users. The image and video generation space has been through this cycle already — early adopters shipped, got sued or notified, and rolled back. Three-D is going to run the same play, and the outputs are more clearly derivative when the training data included copyrighted meshes. Ask your vendor for training data attestation before you commit.
For pure research and prototyping, there's no reason not to grab it. The inference stack is standard PyTorch, the model card is detailed enough to reproduce results, and the community is already publishing Colab notebooks. If you've been waiting for a credible open 3D model to benchmark internal work against, this is it.
The next six months in 3D generation will look a lot like the six months after Stable Diffusion 1.5. Expect a wave of fine-tunes for specific asset categories — game props, architectural elements, character bases — followed by ControlNet-style conditioning modules that let you constrain generation to a reference sketch or silhouette. The closed vendors will respond with either dramatic price cuts or a pivot to higher-end enterprise features. If you're building in this space, the strategic question isn't which model to use today. It's whether you're building on top of the generation layer or trying to compete with it — because the generation layer is now commodity, and the value is going to migrate to whatever sits on top.
A month ago I wrote a blog post about how Hy3 was topping the OpenRouter rankings despite no one talking about it: https://news.ycombinator.com/item?id=48317294As of today, it has fallen to 8/9th on the rankings. I don't see a reason where you would use this model over compe
Novita is offering free Hy3 on OpenRouter until July 21sthttps://openrouter.ai/tencent/hy3:freehttps://x.com/novita_labs/status/2074158304159510819
This is admittedly an aside from the content of the post itself, but... why do so many mobile sites insist on preventing zooming in and seem to share the same incredibly buggy image zooming? It's quite frustrating.
Curious how people feel about this compared to DS4 Flash, given they are pretty close in size. Also curious how well it holds up to heavy quantization.DS4 Flash can currently run reasonably well on systems with ~96gb+ RAM, I wonder if Hy3 can compete there.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Pelican from a few days ago: https://simonwillison.net/2026/Jul/6/hy3/ - I was using the free tier on OpenRouter, which expires on July 21st.I tried the preview model 41 days ago and got a pelican with a "change pelican color" button: https://st