HunyuanImage 3.0

Overview

HunyuanImage 3.0 is Tencent's flagship image-generation model, released on September 28, 2025, whose project News log marks that date as "HunyuanImage-3.0 Open Source." The technical report (arXiv:2509.23951) landed the same day and describes the model as "a native multimodal model that unifies multimodal understanding and generation within an autoregressive framework, with its image generation module publicly available."

The architectural bet is the story. Where Stable Diffusion 3.5 and most of the field build diffusion transformers, HunyuanImage 3.0 generates images autoregressively from a single multimodal Mixture-of-Experts backbone. Tencent's abstract: "we successfully trained a Mixture-of-Experts (MoE) model comprising over 80 billion parameters in total, with 13 billion parameters activated per token during inference, making it the largest and most powerful open-source image generative model to date." The repository puts the expert count at 64.

Two follow-up checkpoints shipped on January 26, 2026: HunyuanImage-3.0-Instruct, described in the News log as a "Release of Instruct (with reasoning) for intelligent prompt enhancement," and HunyuanImage-3.0-Instruct-Distil, a "Distilled checkpoint for efficient deployment." Instruct is what turned the model from a text-to-image system into an image editor — it is tagged image-to-image on Hugging Face.

A note on the word "open source"

HunyuanImage 3.0 is not open source. Tencent calls it that, in the abstract quoted above, and hundreds of downstream articles repeat the phrase. The weights are freely downloadable, but the license they ship under is the Tencent Hunyuan Community License, which is not an OSI-style license. Its LICENSE file states, in capitals:

"THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW."

and defines the Territory as "the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea." A separate clause requires anyone whose products exceed "100 million monthly active users in the preceding calendar month" to "request a license from Tencent, which Tencent may grant to You in its sole discretion."

A license that carves out three of the world's largest economies by geography is open weights under a restricted license, not open source.

The contrast with Tencent's own LLM is the sharpest fact on this page. Tencent Hy3, the company's 295B language model, is tagged apache-2.0 on Hugging Face — a genuine open-source license, with no territorial carve-out. tencent/HunyuanImage-3.0 is tagged license: other, license_name: tencent-hunyuan-community. Same company, same Hunyuan brand, released nine months apart, and materially different rights. Tencent's licensing is per-modality, and you cannot infer one model's terms from another's.

Capabilities

• Autoregressive, not diffusion: Image tokens are generated by the same MoE transformer that reads text, rather than by iterative denoising. Tencent frames this as unifying understanding and generation "within an autoregressive framework."
• Native chain-of-thought before generation: The Instruct card says the model "performs structured thinking to analyze user's input image and prompt" and to expand the user's intent into fuller instructions — a chain-of-thought pass that precedes pixels.
• Prompt self-rewriting: Instruct "automatically enhances sparse or vague prompts into professional-grade, detail-rich descriptions," removing much of the prompt-engineering burden.
• Image editing and multi-image fusion: Adding and removing elements, style changes, background replacement, and combining up to three reference images.
• Sparse activation at 80B scale: 64 experts, ~13B active parameters per token, so serving cost tracks the active slice rather than the 83B checkpoint.
• Distilled fast path: Instruct-Distil is a knowledge-distillation checkpoint that the repository recommends running with --diff-infer-steps 8.

Technical Specifications

• Hugging Face repositories: tencent/HunyuanImage-3.0 (base, text-to-image), tencent/HunyuanImage-3.0-Instruct (image-to-image), tencent/HunyuanImage-3.0-Instruct-Distil
• Architecture: native multimodal autoregressive MoE; model type hunyuan_image_3_moe
• Parameters (model card): 80B total, 13B activated per token, 64 experts
• Parameters (safetensors index): 83,009,199,459 for base and Instruct; 83,044,867,427 for Instruct-Distil. The "80 billion" figure is Tencent's rounded headline, not the tensor count.
• Precision split: 81.75B parameters in BF16, 1.26B in F32 (base repo)
• Checkpoint size: ~169 GB on disk
• GPU memory: "≥ 3 × 80 GB" (base); "≥ 8 × 80 GB" (Instruct and Instruct-Distil)
• Runtime: Python 3.12+, CUDA 12.8; vLLM acceleration added October 30, 2025
• License: Tencent Hunyuan Community License (license: other, license_name: tencent-hunyuan-community)
• Technical report: arXiv:2509.23951, v3 revised June 26, 2026

Tencent does not publish a training-token count, a training-data description, or a resolution ceiling for HunyuanImage 3.0, so this page does not state any. The understanding half of the "native multimodal" model is also not released — the abstract is explicit that only "its image generation module" is public.

Use Cases

• Long, complex prompts: The autoregressive backbone was built to parse instructions of a thousand-plus characters and reason over them before generating.
• Instruction-driven editing: Object insertion and removal, style transfer, and background replacement through HunyuanImage-3.0-Instruct, currently ranked 19th on arena.ai's image-edit board.
• Chinese and English text rendering: Bilingual glyph rendering inside images, a persistent weakness of Western diffusion models.
• Multi-reference composition: Blending up to three source images into one output, for product shots and character consistency.
• Research on unified multimodal generation: One of very few large autoregressive image generators with downloadable weights, and the standard reference point for the architecture.
• On-premise generation outside the excluded territories: Viable where the license permits it and 3×80 GB of GPU memory is available.

Performance / Benchmarks

arena.ai (the former LMArena, rebranded January 2026), leaderboards dated July 5, 2026:

Both boards are led by OpenAI's gpt-image-2 (medium) — 1385±5 on text-to-image, 1466±4 on image edit. The gap is roughly 230 Elo on text-to-image and 164 on image edit. Both Hunyuan entries carry enough votes to bound their ratings tightly, at ±3 Elo. On the text-to-image board only one of the 22 models ranked above hunyuan-image-3.0 has more votes than its 172,762. On the image-edit board, though, 270,964 votes is unremarkable: six of the 18 models ranked above hunyuan-image-3.0-instruct carry more, seedream-4.5 most of all at 880,645.

Note that arena.ai and Artificial Analysis are different organisations with different Elo scales. Scores from one are not comparable to scores from the other, and none are mixed here.

Vendor-reported human evaluation. Tencent's technical report describes a GSB (Good/Same/Bad) study over "1,000 text prompts" scored by "over 100 professional evaluators." It states: "HunyuanImage 3.0 achieves a relative win rate of 14.10% compared to HunyuanImage 2.1," and "HunyuanImage 3.0 achieves relative win rates of 1.17%, 2.64%, and 5.00% compared to Seedream 4.0, Nano Banana, and GPT-Image, respectively."

Read those margins carefully. A 1.17% relative win rate over Seedream 4.0 is a statistical tie, not a lead — and Tencent chose the prompts, ran the study, and reports the result. Tencent also publishes an SSAE (Structured Semantic Alignment Evaluation) comparison over 3,500 key points in 12 categories, but presents it as a figure rather than a numeric table; this page therefore quotes no SSAE numbers.

Limitations

• Not open source, and geographically restricted. The Tencent Hunyuan Community License excludes the European Union, the United Kingdom, and South Korea outright, and caps free use at 100 million MAU. Any page describing the model as "open source" — including Tencent's own abstract — is using the term loosely.
• Datacenter-only hardware. "≥ 3 × 80 GB" for the base model and "≥ 8 × 80 GB" for the Instruct variants. No consumer-GPU checkpoint has been released, and there is no official quantised release. The technical report describes a pruned 20B variant that "can be deployed on a single 24GB RTX 4090 GPU," but those weights are not published.
• Only half the "native multimodal" model is public. Tencent releases "its image generation module." The understanding module is not published, so the unified architecture cannot be fully reproduced.
• Vendor-reported benchmarks. The GSB win rates and SSAE results are Tencent's own, on Tencent's prompt set. The largest margin is against Tencent's previous model; the margins against external competitors are 1–5%.
• Mid-table on independent evaluation. Rank 23 of 72 models on arena.ai's text-to-image board, ~230 Elo behind the leader, on a board updated July 5, 2026.
• Ageing base checkpoint. The base weights date from September 2025 and were last modified January 28, 2026. Competing open-weight image models have shipped since.
• No published USD pricing. Tencent Cloud quotes HY-Image-V3.0 in 元 per image. Third-party USD figures circulating in blog posts are conversions, not a Tencent rate card.

Pricing & Access

Self-hosting is free of charge within the license's limits: no fee, but no EU/UK/South Korea use, and a 100M-MAU ceiling. Weights are at tencent/HunyuanImage-3.0.

Hosted access via Tencent Cloud. The TokenHub 模型价格 ("Model Pricing") page, last updated 最近更新时间：2026-07-07, lists under 图像生成 ("image generation"):

Figures are 元 (CNY) per generated image. Tencent publishes no USD rate card for HunyuanImage, and no conversion is implied here. Tencent does not document which checkpoint HY-Image-V3.0 serves, so the hosted endpoint cannot be equated with the released HunyuanImage-3.0 weights.

Third-party hosting. Both fal.ai (base and Instruct endpoints) and Replicate serve the model. Their prices are set by them, not by Tencent.

Adoption, from the Hugging Face API on July 8, 2026 (downloads is a trailing-30-day count):

The Instruct variant now out-downloads the base model, which is consistent with editing rather than pure text-to-image being the draw.

Ecosystem & Tools

Tencent's wider generative-media line, for context only — none of it shares HunyuanImage's weights:

• Tencent-Hunyuan/HunyuanImage-3.0 on GitHub — inference code, News log, hardware table, vLLM path
• HunyuanImage 3.0 Technical Report — architecture, SSAE, GSB
• hunyuan.tencent.com/image — Tencent's own image playground, named as the official site in the repository
• tencent/Hunyuan3D-2.1 — Tencent's latest open 3D weights, published June 2025 (36,367 downloads/30d, 1,050 likes). Hunyuan3D 3.0's weights are not public: the Hugging Face API returns no Hunyuan3D-3.0 repository under tencent. The 3.0 platform is usable; the weights are not downloadable.
• tencent/HY-World-2.0 — a world model that outputs editable, persistent 3D scenes (meshes, Gaussian splats) rather than video frames. Repository created April 10, 2026, weights staged through May 2026. Note that it, too, ships under a Tencent community license (tencent-hy-world-2.0-community), not an open-source one.
• tencent/HunyuanVideo-1.5 — the video line, published November 18, 2025, and visibly stalled: 2,437 downloads/30d, and rank 34 with 1196±15 over just 5,476 votes on arena.ai's image-to-video board (dated June 23, 2026). Tencent's image models have community traction that its video models do not.
• fal.ai and Replicate — hosted inference for both base and Instruct

Community & Resources

• HunyuanImage-3.0 on Hugging Face — base weights, model card, license file
• HunyuanImage-3.0-Instruct — reasoning and editing checkpoint, January 26, 2026
• HunyuanImage-3.0-Instruct-Distil — distilled, 8-step deployment
• Tencent Hunyuan Community License — read the Territory definition before deploying
• arena.ai leaderboards — text-to-image and image-edit placements
• Tencent Cloud TokenHub 模型价格 — the CNY-per-image rate card
• Tencent open-sources HunyuanImage 3.0 — TechNode — third-party launch coverage, September 28, 2025
• Compare with Tencent Hy3, Seedream 5.0, Stable Diffusion 3.5, Z-Image, Qwen-Image 2.0, and HiDream-O1-Image