Hunyuan3D 2.1 vs 2.5 vs 3.0 vs TripoSR (2026): Which Hunyuan3D Is Actually Open Source?

Here is the answer people are looking for when they type "Hunyuan3D 2.5 download" or "Hunyuan3D 3.0 GitHub": there is nothing to download. Tencent's GitHub organization contains Hunyuan3D-1, Hunyuan3D-2, and Hunyuan3D-2.1 with weights, plus the sibling projects Hunyuan3D-Omni and Hunyuan3D-Part (open, September 2025), HY3D-Bench (a dataset, February 2026), and two 2026 paper-only repos, Buffalo 1.0 and WorldClaw, that ship a README and no weights. There is no Hunyuan3D-2.5, no PolyGen, no 3.0, and no 3.1 repository, and no matching model on Hugging Face. The 2.5 technical report links only to the hosted platform. Issue #111 on the 2.1 repo and issues #227, #232, and #241 on the 2.0 repo all ask for 2.5 or PolyGen weights and, as of September 2026, have no maintainer reply. A claim circulating in search snippets that 2.5 is "Apache 2.0" is wrong on both counts: no Tencent 3D release uses Apache, and there are no 2.5 weights to license.

What is open comes with strings. Hunyuan3D 2.0, 2.1, Omni, and Part all use the Tencent Hunyuan 3D Community License template, which states that it does not apply in the European Union, the United Kingdom, or South Korea, requires a separate license once your product passes 1 million monthly active users, and forbids using outputs to improve any other AI model. Hunyuan3D 1.0 is stricter still, a non-commercial research license. TripoSR is plain MIT: no territory clause, no user cap, no training restriction, commercial use included. So on pure openness the ranking is TripoSR first, Hunyuan3D 2.1 a real but restricted second, and 2.5 and 3.x not in the race at all.

Each Hunyuan3D generation is mainly a geometry upgrade, and the numbers Tencent publishes tell the story. Hunyuan3D 2.1 ships a 3.3B-parameter shape model (Hunyuan3D-DiT) whose meshes the 2.0 report describes as holeless, which is what makes them usable without a repair pass. Hunyuan3D 2.5 replaced that with LATTICE, a shape foundation model whose largest variant reaches 10B parameters, and Tencent's launch post claimed 1024-resolution geometry with ten times more mesh faces than 2.0. Hunyuan3D 3.0 pushed to 1536-cubed geometric resolution with what Tencent calls 3.6B voxel ultra-HD modeling and three times higher precision, and 3.1 kept that resolution while adding more input views.

The catch is that everything past 2.1 is a hosted claim you cannot benchmark yourself, so treat the resolution figures as Tencent's numbers, not independent results. TripoSR sits at the other end: a single feed-forward reconstruction from one image, fast and clean but with less fine detail than any diffusion-based Hunyuan3D and no ability to reason about the unseen side of the object. For geometry, Hunyuan3D 3.x is the winner on paper, 2.1 is the best you can run yourself, and TripoSR is the one you pick for speed rather than fidelity.

PBR arrived in the Hunyuan3D line with 2.1. The 2.0 family painted RGB textures; 2.1 replaced that with a 2B-parameter Hunyuan3D-Paint model that synthesizes albedo, metallic, and roughness maps, and those weights are open. That makes 2.1 the only Hunyuan3D whose PBR pipeline you can run, inspect, retrain, and own. The 2.5 report says its texture stage was upgraded to physically based rendering as well, and on the current platform PBR is a paid option: Tencent Cloud's billing page lists PBR materials as an extra 10 credits on top of the base generation.

TripoSR is the honest weak link here: vertex colors only, with an optional baked-texture UV on export and no material synthesis. If textured, lit-ready output is the point, Hunyuan3D 2.1 is the pick for a pipeline you own and Hunyuan3D 3.x for the best hosted result. We give this round to 2.1 because an open PBR stage you can run beats a hosted one you rent, and because nothing about 3.x's texture quality can be verified outside Tencent's platform.

Input flexibility is where the hosted line pulls furthest ahead. Hunyuan3D 2.1's open weights are image-conditioned only; multi-view input lives in the 2.0 family's 2mv variant, and text-to-3D in the open repos works by chaining a text-to-image model in front of the image stage. Hunyuan3D 3.0, per Tencent's November 2025 press release, accepts text descriptions, images, or sketches, and up to four multi-view images to tighten geometry and materials; 3.1, announced in January 2026, raises that to eight input views. The ComfyUI partner nodes expose the 2 to 4 view option as well.

TripoSR takes one image and nothing else: no text, no sketch, no multi-view. That is the deliberate trade for its speed and size. If your source material is several photos of a real object, or a text brief with no image at all, Hunyuan3D 3.x is the only option on this page that handles it natively, and it wins this round by a wide margin.

The open and hosted lines are priced in different currencies: VRAM versus credits. Hunyuan3D 2.1 needs 10 GB of VRAM for shape generation, 21 GB for texturing, and 29 GB for the full pipeline, which puts a complete PBR result beyond most consumer cards, though a low-VRAM mode exists and the 2.0 family fits shape plus texture in about 16 GB. Hunyuan3D 3.x needs no hardware at all. Tencent's global launch press release states 20 free generations a day on the platform and 200 free credits for the enterprise API; the Tencent Cloud billing page (last updated January 2026, read through a browser since it renders with JavaScript) lists a standard Professional generation at 25 credits, with multi-view, PBR, and face-count control at 10 extra credits each, and postpaid pricing of about 0.02 USD per credit, so a full-option job lands around one US dollar.

TripoSR needs about 6 GB of VRAM, falls back to CPU, and reconstructs in under 0.5 seconds on an A100, with no per-generation cost once it is running. It also runs hosted on this site, in the browser with no GPU, which is a paid product and not part of the free open model; we say so plainly. For the person whose constraint is the GPU they own or the bill at the end of the month, TripoSR wins this round; for the person who has neither a GPU nor time to set one up, Hunyuan3D 3.x on credits and TripoSR hosted here are the two no-hardware routes.

Hunyuan3D 2.1 has the richest open ecosystem in the family: the official README links community ComfyUI and Unity integrations, the repo lists macOS, Windows, and Linux support, the 2.0 family adds mini, multi-view, Fast, and Turbo variants, and the February 2026 HY3D-Bench release even ships an 0.8B small shape baseline. The caveat is momentum: the 2.0 and 2.1 repos have had little activity since late 2025, and Tencent's newer open work went into Omni (controllable generation built on 2.1) and Part (part-level shape generation) rather than a 2.1 successor.

Hunyuan3D 3.0 and 3.1 live inside Tencent's products: the Hunyuan 3D platform and Studio, Tencent Cloud API, and, since February 2026, ComfyUI partner nodes that call the hosted model. That is convenient and officially supported, but it is an API integration, not a local node. TripoSR is the smallest ecosystem here, with a Hugging Face Space, a Gradio app, community ComfyUI nodes, and the hosted generator on this site. For the widest open tooling, 2.1 takes the round; for a supported hosted pipeline, 3.x; for the shortest path to a first mesh, TripoSR.