EliovpAI/MiniMax-H3-Image-FP8-Paiton-RDNA4

MiniMax-H3-Image FP8 for Paiton on RDNA4

FP8 weights of MiniMax-H3-Image for the Paiton native runtime on one AMD Radeon AI PRO R9700 (32 GB). Use them with the container in Eliovp-BV/paiton-vllm-plugin under models/MiniMax-H3-Image, which downloads this repository on its first start.

These are runtime weights for Paiton, not a general FP8 checkpoint. They are stored in our own container formats: the transformer modulation (AdaLN) is precomputed into tables for the fixed 6-step schedule, and the projection layouts are reordered for the Paiton kernels. Only the Paiton image for the R9700 (gfx1201) loads them; diffusers, vLLM-Omni and ComfyUI cannot. That is why the name ends in RDNA4.

MiniMax-H3-Image is MiniMax's 33B image model for text to image and image editing with up to 16 reference images, at 4 to 16 MP and 6 sampling steps. With these weights it runs on a single 32 GB R9700.

Speed

Request	Time
text to image, 2048 x 2048	6.9 s
text to image, 1024 x 1024 (made at 2048 x 2048, then resized)	7.1 s
text to image, 4096 x 4096	33.4 s
edit, 4096 x 4096, one 1024 x 1024 reference	39.9 s
edit, 4736 x 3552, one 4736 x 3552 reference	50.1 s
edit, 4096 x 4096, two references	66.9 s
edit, 4096 x 4096, three references	88.5 s

Server time per request, from the HTTP request to the finished PNG, with a warm server in the Paiton container on 11 Nov 2026. It includes the text encoder, all 6 steps, the VAE decode and PNG encoding. The 2048 and 4096 text-to-image rows are rounded from three runs each (6.87 to 6.99 s and 33.40 to 33.44 s); the other rows are one run each. The card ran at its 300 W power limit and used at most 27.6 GiB of its 32 GB.

Examples

Made with these weights in the Paiton container on a single R9700: 4096 x 4096, 6 steps, about 34 s each, shown here scaled to 2048 x 2048. The prompts ask for things image models often get wrong: a lot of text, a detailed face, and objects in stated positions. We picked the better of two seeds for each.

Concert poster (seed 23)

Prompt

A screen-printed retro concert poster in orange, teal and cream on textured paper. Top line in huge bold condensed letters: "SUMMER SOUND FESTIVAL". Below: "JULY 18 - 20 · RIVERSIDE PARK". A stylized illustration of a sunset over a river with a stage. A lineup list in clean sans-serif, one per line: "THE NORTHERN LIGHTS", "MARA & THE WAVES", "ECHO VALLEY", "DJ KOBALT", "THE PAPER KITES CLUB", "LUNA RAY". At the bottom: "TICKETS FROM 45 · GATES OPEN 14:00". Crisp letterforms, print texture, no misspellings.

Portrait (seed 11)

Prompt

A close-up portrait photograph of an elderly fisherwoman in her seventies on a windy harbor, deep wrinkles, freckles and sun spots, silver hair blowing across her face, light grey-blue eyes, a hand-knitted mustard wool sweater with loose threads, soft overcast daylight, 85 mm lens, shallow depth of field, every pore and strand of hair sharp.

Bookshop window (seed 11)

Prompt

A photograph of the window of an old independent bookshop at dusk, warm light inside. Gold leaf lettering on the glass: "OWL & LANTERN BOOKS" and below it "Rare · Used · New since 1952". In the window, small hand-lettered cards: "Staff Pick of the Week", "Poetry - 20% off", "Signed copies inside", "Open late on Fridays". Stacks of books with readable spine titles: "The Quiet Sea", "A Map of Small Things", "Northern Recipes", "Letters to Clara". Reflections on the glass, sharp detail.

Objects in stated positions (seed 23)

Prompt

A minimal 3D render on a white table: a red cube stacked on top of a blue sphere, a green cylinder standing to the left of them, a yellow cone behind them, and a small purple pyramid in front on the right. Soft studio light, soft shadows, matte materials.

Newspaper front page (seed 23)

Prompt

A photograph of the front page of a broadsheet newspaper lying on a wooden café table next to a cup of espresso, morning light. Masthead in large black blackletter type: "The Harbor Gazette". Under it a thin line: "Tuesday, 4 March 2026 · Price 2.50 · Est. 1887". Main headline in bold serif capitals: "CITY OPENS NEW HARBOR BRIDGE". Subheadline: "Thousands cross on foot as the 2.4 km span opens to traffic". Below, a large black and white photo of a suspension bridge at dawn, then three columns of small readable body text, and a side box titled "WEATHER" with "Sunny, 14°C". Sharp focus, every letter crisp.

Not everything works. Small body text, such as the newspaper columns and some book spines, comes out as lettering that looks right from a distance but does not read as words. Hands in close-up, a clock asked to show three o'clock and a horse asked to ride on an astronaut's back came out wrong in our tries.

What is in it

Path	Contents	Size
dit/	Transformer: FP8 projections with GPTQ rounding, modulation tables for the 6-step schedule	20.3 GB
te/	Qwen3-VL text encoder, decoder layers 0 to 49 in FP8, embedding and vision tower in bf16	27.1 GB
vae/	VAE decoder (int8 blocks) and encoder	0.5 GB
tokenizer/tokenizer.json	MiniMax's tokenizer, unmodified	7 MB

manifest.json lists every file with its size and SHA-256. MODIFICATIONS.md says how each file was made from the original checkpoint at revision 00dd081ebdd33315adf2c4d84f64aa327ddf3039.

Quality

A 30-case gate (24 text-to-image prompts from 4 to 17 MP in several aspect ratios, 6 edits with 1 to 3 references) compared these weights on the R9700 with the BF16 model in diffusers, using the same token ids and the same noise.

	BF16	These weights
CLIP-L text-image score, mean	32.31	32.18
CLIP-L image similarity to the BF16 image, mean / lowest		0.979 / 0.855
DINOv2 image similarity to the BF16 image, mean / lowest		0.945 / 0.699

Prompt alignment matches BF16 within the case-to-case spread. The images differ from the BF16 ones mainly in layout: a few-step sampler follows a different path after small numeric changes. Text in images and faces render correctly. Edits stay closest to the BF16 result (image similarity 0.96 to 0.999). One seed per prompt was scored.

License

MiniMax-H3-Image is licensed under the MiniMax-H3-Image Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved. The full agreement is in LICENSE and applies to these files, which are modified files of MiniMax-H3-Image (see NOTICE and MODIFICATIONS.md). Read it before you download. In short:

• The license does not cover the European Union, the United Kingdom, the Republic of Korea or the United States.
• A commercial product or service that uses the model must show "MiniMax-H3-Image" prominently in its user interface.
• Offering the model to third parties as a hosted service, or annual revenue above US$1 million, needs prior written authorization from MiniMax.
• Users you give access to must be bound by the use restrictions of Section V and the Acceptable Use Policy.
• If you pass these files on, include LICENSE and NOTICE.