AI 3D Model Generation in 2026: Every Major Release, Dated
Last updated: September 2026.
The state of AI 3D model generation in September 2026: Microsoft's TRELLIS.2 (released 16 December 2025, MIT licence) is still the best model you can download and run, Meshy 7 (12 August 2026) and Tripo's H3.1 and P1.0 (announced 9 March 2026) are the newest hosted engines, and Tencent's Hunyuan3D 3.x family, the most capable geometry engine on paper, remains hosted-only with the last open weights at version 2.1 from June 2025. Below is every major release in this field with its date, maker, open-weights status and licence, checked against the GitHub release notes, Hugging Face model cards and vendor announcements on the day this page was written.
This page is a changelog, not a review. We update it monthly (this revision is 6 September 2026) and the roundup of the best AI 3D model generators is where we say which one to use and why. Where a date is an announcement rather than a weights release, the table says so.
Quick answer
- The release that matters most for a game project right now: TRELLIS.2, 16 December 2025. It's a 4B-parameter image-to-3D model under the MIT licence with native PBR materials and GLB export, and it's the only frontier-quality model whose licence has no territory, revenue or user-count clause.
- The newest hosted engines: Meshy 7 (12 August 2026, built around image alignment) and Tripo P1.0 and H3.1 (9 March 2026, native 3D diffusion with a two-second low-poly mode). Both are closed, both need a paid plan to download.
- Best geometry, hosted only: Rodin Gen-2.5 (26 May 2026, ten-million-polygon raw output) and Tencent's Hunyuan 3D 3.1 (28 January 2026 on the global platform, 1536³ geometry).
- The licence to read twice: every Tencent Hunyuan3D open release (2.0, 2.1, Omni, Part) ships under a Community License that does not apply in the EU, UK or South Korea.
- Try two of the open models without a GPU: Cinevva's 3D Model Generator runs Hunyuan3D and TRELLIS.2 in the browser.
Before the full table, here is the shape of the last two years on one line. The open releases cluster in 2025, and 2026 so far has been a hosted year with Pixal3D the single open exception.
Release timeline
Newest first. "Open weights?" means downloadable checkpoints, not just a paper or demo. Dates are from the primary source named in the notes column, and "announced" flags a hosted or press release where no weights shipped.
| Date | Model | Maker | Open weights? | Licence | What changed |
|---|---|---|---|---|---|
| Sep 2026 | Pixal3D (multi-view inference) | Tencent ARC | Yes | MIT | Multi-view inference code added to the May release (GitHub README) |
| 12 Aug 2026 | Meshy 7 | Meshy | No | Hosted, Pro and above to download | New foundation model scored on image-to-3D alignment: 81.0% proportion, 79.7% spatial, 59.8% surface detail on Meshy's benchmark. Ultra Mode single-view only at launch (Meshy blog) |
| Aug 2026 | Meshy 6 Lite | Meshy | No | Hosted | Faster, lighter Meshy 6 variant for draft iteration. Meshy 5 discontinued the same month (Meshy help centre) |
| 7 Aug 2026 | Meshy multiview upgrade | Meshy | No | Hosted | Main view plus up to three auxiliary views drive both geometry and texture (Meshy blog) |
| 22 Jul 2026 | Meshy native 8K textures | Meshy | No | Hosted, paid tiers | Base colour generated natively at 8K by multiview diffusion (Meshy blog) |
| 4 Jun 2026 | Meshy 3D Agent (beta) | Meshy | No | Hosted | Conversational concept-to-model flow with eight export formats (Meshy blog) |
| 26 May 2026 | Rodin Gen-2.5 | Hyper3D (Deemos) | No | Hosted, paid plans | Raw output above 10 million polygons, 4-second drafts up to 80-second tiers, raw tri, quad and Smart Low-Poly outputs (press release) |
| May 2026 | Pixal3D | Tencent ARC | Yes | MIT | Pixel-aligned image-to-3D with PBR textures. Inference code, demo, training code and a TRELLIS.2-backbone variant all in May. SIGGRAPH 2026 paper (GitHub README) |
| 31 Mar 2026 | Rodin Gen-2 Edit ("3D Nano Banana") | Hyper3D | No | Hosted | Select a region on any uploaded mesh and edit it with a text prompt (press release) |
| 9 Mar 2026 | Tripo P1.0 and H3.1 | Tripo (VAST) | No | Hosted, paid to commercialise | Announced for GDC 2026: P1.0 is a native 3D diffusion model producing engine-ready assets "in as little as two seconds", H3.1 is the high-detail flagship. API version string v3.1-20260211 (press release, Tripo API docs) |
| 28 Jan 2026 | Hunyuan 3D 3.1 | Tencent | No | Hosted, 20 free generations a day | Announced on the global platform, hosted only (Tencent Hunyuan on X) |
| 18 Jan 2026 | Meshy 6 (GA) | Meshy | No | Hosted, free tier CC BY 4.0 | Low Poly Mode for game developers, refined hard-surface geometry, multi-colour print export (Meshy blog) |
| 16 Dec 2025 | TRELLIS.2 (4B) | Microsoft | Yes | MIT | O-Voxel sparse representation, full PBR (base colour, roughness, metallic, opacity), GLB textures to 4096², 24 GB VRAM (GitHub release commit, HF card) |
| 26 Nov 2025 | Hunyuan 3D global platform | Tencent | No | Hosted, 20 free generations a day | 3d.hunyuanglobal.com launched worldwide running Hunyuan 3D 3.0 (Tencent newsroom) |
| 19 Nov 2025 | SAM 3D Objects and SAM 3D Body | Meta | Yes | SAM License | Single-photo reconstruction. Objects outputs Gaussian splats, Body outputs a parametric human mesh (GitHub READMEs) |
| 16 Oct 2025 | Meshy 6 Preview | Meshy | No | Hosted | Sculpt-level detail, bulk generation of up to 10 assets, 3D-to-image and video workspace (press release) |
| 1 Oct 2025 | Rodin Gen-2 | Hyper3D | No | Hosted | 10B parameters, recursive part-based generation, claimed 4x mesh quality over Gen-1.5 (press release) |
| 28 Sep 2025 | Hunyuan3D-Part (P3-SAM + X-Part) | Tencent | Yes | Hunyuan 3D-Part Community License | Native 3D part segmentation and part-complete generation. Excludes EU, UK, South Korea (GitHub initial commit, LICENSE) |
| 25 Sep 2025 | Hunyuan3D-Omni | Tencent | Yes | Hunyuan 3D Omni Community License | Control signals: point cloud, voxel, skeleton, bounding box. 10 GB VRAM. Excludes EU, UK, South Korea (GitHub model zoo, License.txt) |
| 16 Sep 2025 | Hunyuan 3D 3.0 | Tencent | No | Hosted, Tencent Cloud API | Announced at Tencent's Global Digital Ecosystem Summit: 1536³ geometry, 3.6 billion voxels, claimed 3x accuracy. Weights never released (Tencent News, Sina) |
| 26 Aug 2025 | Meshy 5 (GA) | Meshy | No | Hosted | Multi-view input, 500+ animations for rigged characters. Preview ran from 28 July 2025 (press releases) |
| Aug 2025 | Tripo 3.0 | Tripo (VAST) | No | Hosted | Ultra Mode up to two million polygons. API version string v3.0-20250812, beta announced 20 August 2025 (Tripo API docs, press release) |
| 13 Jul 2025 | PartCrafter | Peking University et al. | Yes | MIT | Multiple parts or objects from one RGB image in a single pass. NeurIPS 2025 (GitHub README) |
| 23 Jun 2025 | Hunyuan3D 2.5 (technical report) | Tencent | No | Hosted | 10B-parameter engine reported at 1024 geometric resolution. Announced on the hosted platform in April 2025 per press coverage. An open-source request (issue #111) has sat unanswered since July 2025 |
| 13 Jun 2025 | Hunyuan3D 2.1 | Tencent | Yes | Hunyuan 3D 2.1 Community License | 3.3B shape model plus 2B PBR paint model with full training code. 10 GB shape, 21 GB paint, 29 GB both. Excludes EU, UK, South Korea, and 1M MAU cap (GitHub README, LICENSE) |
| 11 Jun 2025 | PartPacker | NVIDIA | Yes | NVIDIA Source Code License (non-commercial) | Part-level generation via dual volume packing, about 10 GB VRAM (arXiv, GitHub, license.md) |
| 30 May 2025 | Direct3D-S2 | DreamTech | Yes | MIT | 1024³ geometry trained on 8 GPUs. 10 GB VRAM at 512, about 24 GB at 1024 (GitHub README) |
| 20 May 2025 | Sparc3D | Math Magic (Hitem3D / Hi3D) | No | Paper and demo only | 1024³ watertight geometry. Weights stayed behind the Hitem3D product, which the repo's own issues complain about (arXiv, GitHub issues) |
| 13 May 2025 | Step1X-3D | StepFun | Yes | Apache 2.0 | 1.3B geometry plus 3.5B texture, 27 to 29 GB VRAM for both stages (GitHub README) |
| 8 May 2025 | Rodin Gen-1.5 | Hyper3D | No | Hosted | 4B+ parameters, quad mesh and PBR direct from image or text (press release) |
| 16 Apr 2025 | UniRig | VAST | Yes | MIT | Autoregressive skeleton prediction plus skinning weights, 8 GB VRAM. SIGGRAPH 2025 (arXiv, GitHub README) |
| Mar to Apr 2025 | Hunyuan3D-2 mini, mv, Turbo, FlashVDM | Tencent | Yes | Hunyuan 3D 2.0 Community License | 0.6B mini shape model, multiview shape model, distilled turbo shape and paint models (GitHub README news) |
| 28 Mar 2025 | Hi3DGen (Stable3DGen) | Stable-X | Yes | MIT | Normal-bridged geometry on a TRELLIS backbone, checkpoint trellis-normal-v0-1 (arXiv, GitHub, HF) |
| 25 Mar 2025 | TRELLIS text models and training code | Microsoft | Yes | MIT | TRELLIS-text base, large and xlarge (342M to 2.0B) (GitHub README news) |
| Mar 2025 | TripoSG | VAST | Yes | MIT | 1.5B rectified-flow shape model, 8 GB VRAM (GitHub README) |
| Jan 2025 | Tripo 2.5 | Tripo (VAST) | No | Hosted | Multi-image input, Blender and Unity plugins (Tripo blog) |
| 21 Jan 2025 | Hunyuan3D 2.0 | Tencent | Yes | Hunyuan 3D 2.0 Community License | Shape and paint models on Hugging Face. 6 GB VRAM shape only, 16 GB with texture. Excludes EU, UK, South Korea (GitHub README, LICENSE) |
| 8 Jan 2025 | SPAR3D | Stability AI | Yes | Stability AI Community License | Point-aware mesh in 0.7 seconds, editable point cloud, free under $1M annual revenue (Stability announcement, LICENSE.md) |
| 2 Dec 2024 | TRELLIS | Microsoft | Yes | MIT | Structured latents, 1.2B image model, multi-image conditioning added 18 Dec, 16 GB VRAM (arXiv, GitHub README) |
| 19 Sep 2024 | Tripo 2.0 | Tripo (VAST) | No | Hosted | API version string v2.0-20240919, PBR materials (Tripo API docs) |
| 1 Aug 2024 | Stable Fast 3D | Stability AI | Yes | Stability AI Community License | UV-unwrapped mesh in 0.5 seconds, about 6 to 7 GB VRAM (Stability announcement, GitHub README) |
| May 2024 | Rodin Gen-1 | Hyper3D (Deemos) | No | Hosted | 1.5B native 3D model, quad faces and PBR in about 40 seconds, open trials (Deemos on X) |
Q4 2025: the open-weights ceiling moved
The quarter had three releases that changed what an indie could run on their own hardware. TRELLIS.2 landed on 16 December 2025 with a 4B-parameter model, an O-Voxel representation that handles open surfaces and internal structure, and PBR textures to 4096², all under MIT. The catch is the 24 GB VRAM requirement, which puts it on a 3090, 4090 or a rented A100 rather than a laptop.
Meta's SAM 3D pair on 19 November 2025 solved a different problem: reconstruction of a real object from one photo, with occlusion, outputting Gaussian splats rather than a mesh. The SAM License allows commercial use with the usual Meta exclusions around military end uses. For a game pipeline it's a reference capture tool, not a prop generator.
Tencent spent the quarter going the other way. Hunyuan3D-Omni (25 September) and Hunyuan3D-Part (28 September) shipped weights for controllable generation and native part decomposition, both useful for game assets where you want a door that's a separate mesh from the wall. But the 3.0 engine announced on 16 September and the global platform that launched on 26 November stayed hosted, and the open releases carry the Community License that does not apply in the EU, UK or South Korea. If your studio or your players are in those territories, that clause decides the question for you.
On the hosted side, Rodin Gen-2 (1 October) introduced recursive part-based generation at 10B parameters, and Meshy 6 Preview (16 October) pushed sculpt-level detail. Both were previews of where 2026 would go: bigger models, more part structure, and download rights tied to a subscription.
Q1 2026: hosted engines went production-grade
Three vendors shipped their "for pipelines, not demos" release within ten weeks of each other. Meshy 6 went GA on 18 January with a Low Poly Mode aimed at game developers. Tencent put Hunyuan 3D 3.1 on the global platform on 28 January, hosted only, twenty free generations a day. Tripo announced P1.0 and H3.1 on 9 March for GDC, and P1.0 is the interesting one: a native 3D diffusion architecture that resolves the whole object at once and produces low-poly, engine-ready output in about two seconds. That is a different bet from the "generate dense, then remesh" pattern everyone else uses. Tripo's API had already been serving a v3.1-20260211 version string a month before the announcement.
For topology this was the quarter the vendors stopped pretending. Meshy's Low Poly Mode, Tripo's P1.0, and Rodin's quad options all exist because raw diffusion output has no edge loops and everyone knows it. None of them produce the concentrated loops a face or shoulder needs for deformation. They produce evenly distributed quads or a clean low-poly shell, which is exactly right for props and exactly not enough for a hero character.
Nothing open of comparable weight shipped this quarter. The gap between the hosted 3.x-class engines and the open 2.x-class weights widened.
Q2 2026: detail on the hosted side, alignment on the open side
Rodin Gen-2.5 on 26 May 2026 is the high-water mark for raw geometric detail: more than 10 million polygons in the top tier, four-second drafts at the bottom, and three output types (raw triangles, quads, Smart Low-Poly). Two months earlier, on 31 March, Hyper3D shipped Gen-2 Edit, which lets you mask a region of an existing mesh and change it with a prompt. Region editing is the feature that turns a generator into an iteration tool, and it's still rare.
Meshy launched its 3D Agent beta on 4 June, a chat flow from concept through reference image to downloadable model. It's the same two-step pattern (image model first, image-to-3D second) that every serious pipeline already uses, packaged for people who don't know that.
The open release of the quarter was Pixal3D from Tencent ARC in May 2026, under MIT: a pixel-aligned image-to-3D method that back-projects image features into 3D for reconstruction-level fidelity, with training code and a TRELLIS.2-backbone variant in the same month. It's the first significant open model built on top of TRELLIS.2 rather than alongside it, which is how open ecosystems compound.
Q3 2026: Meshy's texture and alignment quarter
Meshy shipped four releases in seven weeks: native 8K base-colour textures on 22 July (paid tiers), a multiview upgrade on 7 August that uses up to four reference views for both geometry and texture, then Meshy 7 and Meshy 6 Lite in August. Meshy 7 is the one to understand. Instead of claiming "better", Meshy published alignment scores against held-out reference models: 81.0% overall proportion, 79.7% spatial distribution, 59.8% surface details. That surface-detail number is the honest one. Proportion and layout are close to solved for single objects, fine detail isn't, and only Pro and above can download Meshy 7 output.
For PBR textures, this quarter made 8K the marketing number. For a browser game, it's irrelevant: you'll ship 1K or 2K compressed to KTX2, and the win is that downsampling from a clean 8K source gives you fewer baked-in lighting artefacts than a 2K source did. Bake-in in the albedo is still the most common texture defect across every tool on this page.
On the open side, Pixal3D added multi-view inference in September. No new frontier-scale open weights shipped in Q3, and none of the hosted vendors released a model. The 2.1-era Hunyuan weights remain the newest open Tencent shape model, fifteen months on.
What you can run yourself
Eight of the models above have downloadable weights and a licence that permits commercial use, at least somewhere. VRAM figures are from each project's own README.
| Model | Licence | Commercial use | VRAM (README) | Output |
|---|---|---|---|---|
| TRELLIS.2 | MIT | Yes, no conditions | 24 GB minimum | GLB with PBR to 4096², OBJ, PLY, splats |
| TRELLIS | MIT | Yes | 16 GB minimum | Mesh, radiance field, splats |
| Pixal3D | MIT | Yes | Not stated, low-VRAM mode available | Mesh with PBR |
| Hunyuan3D 2.1 | Hunyuan 3D 2.1 Community License | Yes, outside EU, UK, South Korea, under 1M MAU | 10 GB shape, 21 GB paint, 29 GB both | Mesh with PBR |
| Hunyuan3D 2.0 | Hunyuan 3D 2.0 Community License | Same territory and MAU terms | 6 GB shape, 16 GB with texture | Mesh with RGB texture |
| Hunyuan3D-Omni / Part | Hunyuan Community Licenses | Same territory terms | 10 GB (Omni) | Controlled shape / parts |
| Direct3D-S2 | MIT | Yes | 10 GB at 512, about 24 GB at 1024 | OBJ mesh |
| Step1X-3D | Apache 2.0 | Yes | 27 to 29 GB both stages | Watertight mesh plus texture |
| Hi3DGen | MIT | Yes | Not stated | Geometry |
| TripoSG | MIT | Yes | 8 GB | Shape |
| SPAR3D | Stability AI Community License | Yes under $1M annual revenue | About 6 GB, 10.5 GB default | GLB |
| Stable Fast 3D | Stability AI Community License | Yes under $1M annual revenue | About 6 to 7 GB | GLB |
| SAM 3D Objects | SAM License | Yes, with Meta's end-use exclusions | Not stated | Gaussian splat (.ply) |
| PartCrafter | MIT | Yes | Not stated | Part-decomposed mesh |
| UniRig | MIT | Yes | 8 GB | Skeleton and skin weights |
| PartPacker | NVIDIA Source Code License | No, research only | About 10 GB | Part-decomposed mesh |
The same numbers as a chart, sorted by the memory you need for the full model, with the licence terms as the colour.
Three things the table can't show. First, "open weights" for Hunyuan3D 2.1 includes full training code, which none of the MIT models offered until TRELLIS.2 and Pixal3D shipped theirs, so it's still the model to fine-tune on your own art style if the territory clause doesn't rule you out. Second, Sparc3D is not on this list because its weights were never released, despite the GitHub repo and the Hugging Face demos. Third, if you don't own a 24 GB card, the practical route is renting inference by the second, or using a hosted front end for one of these models.
What this means for a game project
The pipeline for a static prop is settled. Generate a reference image, feed it to an image-to-3D model, check the triangle count in a GLB viewer, decimate and compress with the GLB optimizer, and ship. Every hosted tool and every open model above produces something usable for a rock, a crate, a lamp or a stylised item after that pass. The choice between them is licence and cost, not quality.
Cinevva's 3D Model Generator is the way to try the two open models that matter without installing anything. The tool page describes it as a workbench that generates reference images from a text prompt (or takes an uploaded image or URL), reconstructs the one you pick into a textured mesh, then lets you rig, animate, inspect and export it as GLB without moving files between services. Generating and previewing uses the credits shown in the workbench, your first GLB export is free, and further exports come with the Standard and Pro plans. The mesh engine picker on that workbench offers Hunyuan3D and TRELLIS.2, so you can compare the Tencent and Microsoft models on the same reference image, and for characters the character option keeps the references in a riggable pose (deep link: /tools/hunyuan3d#prompt=<text>&character=1).
For characters, the release that matters isn't a generator at all. UniRig (April 2025, MIT) showed that skeleton prediction and skinning could be a learned step, and every hosted vendor now bundles auto-rigging. Cinevva's Auto Rigger takes a GLB, FBX or OBJ up to 50 MB, rigs humanoids on its Fast engine, and returns a GLB with up to six baked animations, with the first rigged character free to download. Standard and Pro remove the cap and add the Pro engine for quadrupeds. If you rig often in Blender, the Blender add-on runs the same engine locally with no per-rig charge for a one-time $29 at the launch price. Whichever rigger you use, the topology caveat from Q1 applies: auto-rigged AI meshes hold up as background characters, and close-up hero characters still need a retopology pass first.
On licences, the one-line version: TRELLIS.2 and Pixal3D output has no strings from the model side, Hunyuan3D 2.x output is fine outside the EU, UK and South Korea, Stability's models are fine under $1M revenue, PartPacker output can't ship at all, and every hosted tool ties download or commercial rights to a paid plan. Read the game asset licences guide before you build a pipeline around any of them.
Common Questions
What's the latest news in AI 3D asset generation in 2026?
The newest releases are Meshy 7 and Meshy 6 Lite (August 2026), Rodin Gen-2.5 (26 May 2026), Tripo P1.0 and H3.1 (announced 9 March 2026), Tencent's Hunyuan 3D 3.1 on the global platform (28 January 2026), and Meshy 6 GA (18 January 2026). The most significant open release of 2026 is Pixal3D (May 2026, MIT), built on the TRELLIS.2 backbone.
When was Hunyuan3D 3.0 released, and is it open source?
Tencent announced Hunyuan 3D 3.0 on 16 September 2025 at its Global Digital Ecosystem Summit and launched it worldwide on the hosted platform at 3d.hunyuanglobal.com on 26 November 2025, followed by 3.1 on 28 January 2026. Neither 3.0 nor 3.1 has released weights. The newest open Hunyuan3D shape model is still 2.1 from 13 June 2025.
What is the TRELLIS 2 release date and licence?
Microsoft released TRELLIS.2 on 16 December 2025 (the "Init Release" commit and the Hugging Face paper date agree). It's a 4B-parameter image-to-3D model under the MIT licence, needs an NVIDIA GPU with at least 24 GB of memory, and exports GLB with PBR textures up to 4096².
When did Meshy 6 and Meshy 7 come out?
Meshy 6 launched as a preview on 16 October 2025 and reached general availability on 18 January 2026. Meshy 7 followed on 12 August 2026, alongside a lighter Meshy 6 Lite, and Meshy 5 was discontinued the same month. Meshy 7 output can only be downloaded on Pro and above.
When did Tripo 3 and Rodin Gen-2 release?
Tripo's API served version v3.0-20250812 from 12 August 2025, with the 3.0 beta announced on 20 August 2025, and v3.1-20260211 from 11 February 2026 ahead of the P1.0 and H3.1 announcement on 9 March 2026. Rodin Gen-2 was announced by Hyper3D on 1 October 2025 and Gen-2.5 on 26 May 2026.
Is there an open source text-to-3D generator?
Yes. TRELLIS.2 (MIT), Pixal3D (MIT), Hunyuan3D 2.1 (Community License, excludes the EU, UK and South Korea), Direct3D-S2 (MIT) and Step1X-3D (Apache 2.0) all have downloadable weights. Most are image-to-3D, so text-to-3D means generating a reference image first, which is what the hosted tools do internally too. TRELLIS also ships dedicated text-conditioned models.
What is Sparc3D and can I download it?
Sparc3D is a May 2025 paper from the team behind Hitem3D (now Hi3D) describing 1024³ watertight geometry generation. The GitHub repo and several Hugging Face demos exist, but the trained weights were never published and stayed behind the commercial product, which the repo's own open issues complain about.
Related
- Best AI 3D Model Generators in 2026: the roundup, which tool to pick and why
- Best AI Asset Generators for Games: the full AI asset pipeline beyond meshes
- Best Open-Source Generative AI Models for Games: the open model picture across image, audio and 3D
- Game Asset Licenses Explained: what each licence in the table lets you ship
- Free Character Animations and Auto-Rigging: what to do with the mesh once you have it
- Best Free 3D Model Sites for Games: when searching beats generating