whatmodel.sh / image / HunyuanImage 3.0-Instruct

HunyuanImage 3.0-Instruct

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

HunyuanImage 3.0-Instruct is an open-weights image model from Tencent, released January 2026. Its published size is 80B parameters. Licence: tencent-hunyuan-community.

Model information sources

Release date ↗ · Parameter count ↗

MoE: 80B total, 13B activated per token (whole MoE transformer). Same convention as the catalogued hunyuanimage-3-0 (params_b 80); active 13B could go in active_b.

62.2±3.1#32 of 144 image modelsfrom 2 benchmark results across 2 sources* thinly tested, treat the rank as provisional

Score within the LMArena text-to-image comparison set only.

$ how it compares

The image models ranked either side of HunyuanImage 3.0-Instruct, out of 197 with a score, then more from Tencent.

#modelscorewhereRAMreleased
36FLUX.2 [pro] Black Forest Labs63±2api $0.030/image2025-11
37FLUX.2 [flex] Black Forest Labs62±2api $0.050/image2025-11
38HunyuanImage 3.0-Instruct Tencent62±3api $0.090/image +local~203GB2026-01
39Nano Banana (Gemini 2.5 Flash Image) Google62±1api $0.039/image2025-08
40Wan 2.7 Image Alibaba62±2api $0.030/image2026-04
more from Tencent
65HunyuanImage 2.1 Tencent58±2api $0.10/image +local~64GB2025-09
68HunyuanImage 3.0 Tencent57±2api $0.080/image +local~49GB2025-09

see all image models · find the ones that fit your machine

$ what it's good at

image to image / editing▁▁·▃▂▃█▃▃▄▁▁ #22 of 62 * · 54% of its score · 1 test

Beats Seedream 4.5, FLUX.2 [max], Wan 2.7 Image Pro here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
Arena image edit1302.7 ✎▂▁▁▂▆▇█▄▆▅▁▂ beats 31/53avg 1153▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 54%typical
other tests1 not in a task · 42% of its score
testresultvs the others in its rundifficultycountswhy
InfoEdit infographic editing (average success rate)3.2% ✎▄··▄█····▄▄▄ beats 1/6very hard · avg 1.4%▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 42%sharp · too hard for it

$ where to run it

Through an API

api · fal.ai · $0.090/image · fal.ai ↗ · price sourceUSD for one 1-megapixel output: fal bills this endpoint $0.09 per megapixel (price = rate × 1 MP; fal's rounding for 1024×1024 ≈ 1.05 MP is not stated; larger outputs cost proportionally more). Product: fal.ai serverless text-to-image. API model: fal-ai/hunyuan-image/v3/instruct/text-to-image ('Instruct version of Hunyuan-Image 3.0'); the edit endpoint fal-ai/hunyuan-image/v3/instruct/edit has the same $0.09/MP rate. Settings: 1 MP output, one image. Price basis: published per-megapixel list rate × 1 MP. Currency: USD. Not a Tencent direct tariff. fal.ai official model page, observed 2026-09-26.

Other ways to run it

  1. Run it as published · needs ~202.7 GB, NVIDIA only · from huggingface.co/tencent (official)
    git clone https://github.com/Tencent-Hunyuan/HunyuanImage-3.0.git && cd HunyuanImage-3.0 && pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128 && pip install -r requirements.txt && hf download tencent/HunyuanImage-3.0-Instruct --local-dir ./HunyuanImage-3-Instruct && python3 -c "from transformers import AutoModelForCausalLM; m=AutoModelForCausalLM.from_pretrained('./HunyuanImage-3-Instruct', attn_implementation='sdpa', trust_remote_code=True, torch_dtype='auto', device_map='auto', moe_impl='eager', moe_drop_tokens=True); m.load_tokenizer('./HunyuanImage-3-Instruct'); cot,s=m.generate_image(prompt='Make a new fridge magnet from the logo in image 1 using the fridge-magnet material in image 2', image=['./assets/demo_instruct_imgs/input_1_0.png','./assets/demo_instruct_imgs/input_1_1.png'], seed=42, image_size='auto', use_system_prompt='en_unified', bot_task='think_recaption', infer_align_image_size=True, diff_infer_steps=50, verbose=2); s[0].save('image_edit.png')"

Official weights and documentation: huggingface.co/tencent ↗

$ how the score was worked out · LMArena leaderboard dataset 54%, InfoEdit: Probing Global Layo… 42%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
LMArena leaderboard dataset1▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 54%CC BY 4.0
InfoEdit: Probing Global Layout Reasoning in Infographic Editing (arXiv:2609.33286v1)1▇▇▇▇▇▇▇▇▇▇▇▇▇ 42%CC-BY-4.0
prior · pull toward the middle▇ 4%
weighted average= 62.2

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json