Open Source
Open Source

16 Megapixels, No Upscaler: What Does the Repo Actually Ship?

MiniMax H3-Image is an open-weight image generation model whose official repository MiniMax-AI/MiniMax-H3-Image (hyphenated GitHub organization) holds, as verified on November 11, 2026, no runnable inference code - only a model card, a license and resource files, with language unset and a single commit. The README model card states 33.1B parameters in a dense single-stream transformer, distilled to 6 sampling steps (6 NFEs) with no classifier-free guidance, about 16 megapixels at every aspect ratio directly and without an upscaler, BF16 DiT weights at 68.65 GB, and a Qwen3-VL text encoder kept outside the repository; the official launch post separately claims a 4K (4096 x 4096) image in under 10 seconds, Standard 1.5-2x and Highspeed 3-4x speeds, pricing at 50-70% of mainstream offerings and a launch-week 50% off for 4K-and-above requests - all launch-post wording, with no seconds figure anywhere in the README. The API guide caps output area at 20,000,000 pixels, allows 1 to 16 reference images and 1 to 10 images per request, and both endpoints are synchronous. The license stays a one-sentence boundary: community license with GitHub NOASSERTION, terms per the in-repo LICENSE, angled against the October 24 M3.1-Flash deployment SOP.

Published November 12, 20269 min read
<!-- minimax-h3-image-opensource-resource | open-source | 16 Megapixels, No Upscaler: What Does the Repo Actually Ship? -->

16 Megapixels, No Upscaler: What Does the Repo Actually Ship? That is not a rhetorical flourish. It is the most direct impression left after opening the official repository for MiniMax H3-Image: an open-weight image model that pushes output straight to about 16 megapixels and a 4096-class resolution, paired with a repository that holds no runnable inference implementation at all, only a model card, a license file, and a few asset folders. This piece is a single-product teardown, not a leaderboard, not a benchmark comparison, and not a line-by-line deployment guide. What it sets out to answer is narrower: what exactly has been made public in this repository, and which numbers belong to which layer of source material. Every fact here is as of November 11, 2026, checked against four directly reachable first-party sources: the README model card in the official repository, the LICENSE text in that same repository, the official launch blog post, and the official API documentation. Keeping those layers apart matters far more than compressing them into one sentence.

One repository, two ways to read it

Begin with the repository itself. Its exact full name is MiniMax-AI/MiniMax-H3-Image, with a hyphenated GitHub organization name, and it is reachable through this repository link. As of November 11, 2026, the repository carries an empty language field and exactly 1 commit; there is no source directory and no executable inference script inside it. From the contents view, the layout is roughly this:

text
MiniMax-H3-Image/
  LICENSE
  README.md
  assets/
  examples/
  skills/

So this is a model card plus license mirror repository: what is public is the weight documentation and the authorization text, not a full inference codebase. The README does name SGLang, diffusers, vLLM, and ComfyUI as inference frameworks, but those are recommendations and integration notes; the frameworks themselves do not live here. A separate article on comfyui-resource covers that class of inference framework. If you want the model card and the license text directly, the plain address is https://github.com/MiniMax-AI/MiniMax-H3-Image, and the raw channel returns both documents in full. Assuming that open-weight automatically means the inference code is shipped too is the first and most common misreading of a repository like this one.

The model card layer: architecture and weights

Read the README model card as the first layer of facts. It gives architecture and weights, not marketing language. The model is a 33.1B-parameter dense single-stream transformer, with generation and editing sharing one checkpoint. Sampling has been distilled to 6 sampling steps (6 NFEs) with no classifier-free guidance. A hard constraint travels with that: num_inference_steps accepts only an omitted value or 6, and any other value returns HTTP 400. Those three items interlock. The parameter count fixes the model's size, the 6-step schedule compresses sampling to a very short run, and dropping guidance removes an extra forward path. Nothing here is a speed or price claim; it is a description of how the model is built and sampled.

Resolution is the single line in this layer most worth pulling out. The card says about 16 megapixels at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9), generated directly and without an upscaler. The 16 megapixels and the 4096 are two faces of the same 4K target: the 1:1 goal for 4K is exactly 4096 x 4096. Note the card's literal wording is 16 megapixels or 16 MP, with a space between number and unit, and the weight figure 68.65 GB carries a space the same way.

The weights table gives BF16 DiT at 68.65 GB, with sub-items such as transformer 66.28 GB, vae 2.34 GB, and pdd_fused_heads 12 MB; an optional PE model is NVFP4 at 17.9 GB. The text encoder is Qwen3-VL, and the card states plainly that it is not in this repository: model_index.json loads it from the H3 family weights under the text_encoder folder. It is therefore just a name here, with no version suffix attached. The editing endpoint on the service side is /v1/images/edits, and generation and editing run on the same weights. The card also says supports up to 16 reference images as input; this 16 is the ceiling on reference images supplied to a request, not the number of finished images the model returns, a distinction the parameter layer makes explicit. Taken together, the model card answers what the model is, how large its weights are, and how it is loaded. It does not answer what it costs or how many times faster it runs than something else.

The launch post layer: speed and pricing

The second layer comes from the official launch blog post. It is not the same document as the model card, and its framing is different, so every number in this section should be read with the qualifier that it is the official launch post's wording. The post puts 4K image generation into a 10-second era and states that a 4K image, that is, a 4096 x 4096 image, can be generated in under 10 seconds. That time is tied to the 4K 1:1 target only; it is not a claim that every resolution lands inside 10 seconds, and it is not a model card fact. The README contains no seconds or under 10 phrasing anywhere in its text. In other words, the 10-second figure is a launch-post statement, while 6-step distillation is the model card's sampling fact. They cannot be folded into one sentence as though they came from one source.

Speed appears in the launch post in two places, and neither cancels the other. The Speed section says at up to 5x the speed of mainstream alternatives, an upper-bound phrasing. In the Pricing section the framing splits into two tiers: Standard mode is 1.5-2x as fast as mainstream alternatives, while Highspeed mode reaches 3-4x. Both wordings stand, so any reading has to carry its section qualifier and cannot lift a single number out as the only conclusion. The same Pricing section puts pricing at 50-70% of comparable mainstream offerings. For launch week, the post offers 50% off the standard price for image generation requests at 4K resolution and above, but it says only For launch week and gives no specific end date; that detail the official post does not state, and it should not be inferred. The channels listed in the post are Hugging Face, MiniMax API Platform, MiniMax Code, and Hailuo AI, and those are places to use the model, not contents of the repository. The post also carries a vendor-stated superlative about prompt adherence among open-weight models; that is a self-reported claim rather than an independently verified result, and this article neither treats it as an objective conclusion nor uses it to rank anything.

The parameter layer: API documentation boundaries

The third layer is the official API documentation, which gives the boundaries on the calling side. Up to 16 reference images is written there as Reference images 1 to 16, meaning a single request may carry 1 to 16 reference images, while Images per request 1 to 10, default 1 governs how many finished images come back. The two numbers have to be read separately: 16 is the input-side ceiling on reference images, and 10 is the output-side ceiling on generated images. Treating 16 as a count of output images is the single most common slip in paraphrases of this model.

Output area is capped at 20,000,000 pixels, with area tiers of 1K, 2K, and 4K, defaulting to 4K, and the 1:1 target for 4K is 4096x4096. The area has to fall between 655,360 and 20,000,000 pixels and the aspect ratio between 1:3 and 3:1, otherwise the request returns 400. Two model ids are given, MiniMax-H3-Image and MiniMax-H3-Image-Highspeed, matching the Standard and Highspeed tiers from the launch post. The generation endpoint is POST /v1/images/generations and the editing endpoint is POST /v1/images/edits, and both endpoints are synchronous. Prompt length is capped at 32000 characters, and with prompt enhancement disabled only the first roughly 4096 tokens take effect. Reference images are limited to JPG, JPEG, PNG, WEBP, HEIC, and HEIF, with a single file no larger than 64 MiB, side lengths in the 128 to 8192 px range, and a long-to-short side ratio no greater than 4:1. Preview images, masked edits, and transparent backgrounds are listed as not supported. These are hard constraints from the documentation, on a different plane from the launch post's speed and pricing wording, and mixing the two is how a parameter gets mistaken for a promise.

License boundary and the in-site angle

The license gets one sentence of boundary here and no more. The launch post's wording for the model is open-weight, the authorization text in the official repository is the MiniMax-H3-Image COMMUNITY LICENSE AGREEMENT, a community license, and GitHub records license.spdx_id as NOASSERTION. The official blog post does not mention the license name or its commercial conditions, so the LICENSE text in the official repository is the authority for those. Set that boundary against the in-site article minimax-m3-1-flash-openweight-resource from October 24, 2026, and the difference in angle is clear: that piece is about a text model's weight release and deployment steps, while this one stays on an image model's model card facts. One is text, the other image; one is a deployment procedure, the other a model card teardown.

Why the layers are kept apart

Splitting the four sources into a model card layer, a launch post layer, and a parameter layer, plus one sentence of license boundary, is a way of putting every number where it belongs. 33.1B, distillation to 6 steps with no CFG, direct 16 megapixels without an upscaler, 68.65 GB of weights, the Qwen3-VL text encoder, and up to 16 reference images belong to the model card and describe the model itself. The 10-second 4K figure, the up to 5x speed, the two Standard and Highspeed tiers at 1.5-2x and 3-4x, pricing at 50-70%, and the launch-week 50% discount belong to the official launch post, are market-facing wordings, and carry their own section qualifiers. Area, aspect ratio, endpoints, images per request, and reference image formats belong to the API documentation and are hard constraints at call time. Beyond the three layers, the shape of the repository adds one more reminder: what is public here is documentation and authorization, not an inference implementation you can simply run.

For the same reason, when reading an open-weight model like this, aligning the provenance is more useful than rushing to place it on a scale. Ask which sentence comes from the model card, which from the launch post, which from the API documentation, and which details the official material does not state at all. The launch-week discount has no published end date, the 10-second figure is bound to 4K at 4096 x 4096, 16 is a reference image ceiling rather than an output count, and the two speed wordings need their section qualifiers. Those are precisely the points that paraphrases flatten. Separate them and what remains is the set of facts that can be quoted without hedging. Earlier in-site articles on image models, for example qwen-image-2-1-open-source-resource and flux-3-image-klein-opensource-resource, cover their own models and architectures; this piece does not repeat that general architecture discussion, and instead pins down the numbers and wordings specific to H3-Image.

FAQ

Q1: Does this repository contain inference code I can run directly? A1: No. As of November 11, 2026, MiniMax-AI/MiniMax-H3-Image has an empty language field and 1 commit, and its contents are LICENSE, README.md, and asset folders. It is a model card plus license mirror repository. The README recommends SGLang, diffusers, vLLM, and ComfyUI, but those are integration notes, and the inference implementation is not in the repository.

Q2: Is the 16 megapixels output produced by upscaling? A2: No. The README says about 16 megapixels at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9), directly and without an upscaler. The 16 megapixels and the 4096 are two faces of the same 4K 1:1 target, generated directly rather than enlarged afterward.

Q3: Is the 10-second 4K figure written in the model card? A3: No. The README has no seconds or under 10 phrasing. The 10-second figure comes from the official launch post and is limited to a 4K, that is a 4096 x 4096, image, not every resolution. The model card's sampling fact is distillation to 6 steps with no classifier-free guidance; the two come from different sources and cannot be treated as one sentence.

Q4: Does the 16 ceiling refer to reference images or to images returned per request? A4: It refers to reference images. 16 is the input-side ceiling on reference images: the README says supports up to 16 reference images, and the API documentation says Reference images 1 to 16. The finished images returned per request are 1 to 10, default 1. Treating 16 as an output count is the most common misreading.

Q5: On what date does the launch-week 50% discount end? A5: The official launch post says only For launch week, offering 50% off the standard price for image generation requests at 4K resolution and above, with no specific end date; the official post does not state one, so it should not be inferred. The channels listed there are Hugging Face, MiniMax API Platform, MiniMax Code, and Hailuo AI.

Related reading

If you want to widen the image-generation thread, ai-image-generation-comparison-review takes the comparison angle, whereas this piece deliberately avoids ranking and instead unpacks one model's model card and launch post wordings; the two read well side by side.

This article is AI-assisted and human-edited. Last updated: 2026-11-12

FAQ

Does this repository contain inference code I can run directly?
No. As of November 11, 2026, MiniMax-AI/MiniMax-H3-Image has an empty language field and 1 commit, and its contents are LICENSE, README.md, and asset folders. It is a model card plus license mirror repository. The README recommends SGLang, diffusers, vLLM, and ComfyUI, but those are integration notes, and the inference implementation is not in the repository.
Is the 16 megapixels output produced by upscaling?
No. The README says about 16 megapixels at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9), directly and without an upscaler. The 16 megapixels and the 4096 are two faces of the same 4K 1:1 target, generated directly rather than enlarged afterward.
Is the 10-second 4K figure written in the model card?
No. The README has no seconds or under 10 phrasing. The 10-second figure comes from the official launch post and is limited to a 4K, that is a 4096 x 4096, image, not every resolution. The model card's sampling fact is distillation to 6 steps with no classifier-free guidance; the two come from different sources and cannot be treated as one sentence.
Does the 16 ceiling refer to reference images or to images returned per request?
It refers to reference images. 16 is the input-side ceiling on reference images: the README says supports up to 16 reference images, and the API documentation says Reference images 1 to 16. The finished images returned per request are 1 to 10, default 1. Treating 16 as an output count is the most common misreading.
On what date does the launch-week 50% discount end?
The official launch post says only For launch week, offering 50% off the standard price for image generation requests at 4K resolution and above, with no specific end date; the official post does not state one, so it should not be inferred. The channels listed there are Hugging Face, MiniMax API Platform, MiniMax Code, and Hailuo AI.

Related

Open Source

Qwen-Image 2.1 Open Weights Are Back, But Apache Is Not

Alibaba's Qwen team open-sourced Qwen-Image-2.1 on September 20, 2026: text-to-image, image editing and native RGBA transparency in one 7B checkpoint - a 32-layer single-stream DiT (mixed-granularity attention plus prefix KV reuse) + a Qwen3-VL 8B text encoder (which also encodes reference images) + a 64-channel RGBA VAE. About 33GB in BF16, compressible to roughly 14GB via Comfy-Org quantized repacks. Capabilities: native 2048x2048 (seven ratios up to 2752x1536), 40-step inference, up to 10 reference images per call, circle/mask local edits, direct transparent-PNG output (official prompt phrasing provided), plus two 9B prompt rewriters. Reported throughput: vLLM-Omni at 1024 squared, 40 steps, BF16 peaks at 34GB and 3.3-4.5s per image (single source); Alibaba says a 3090 runs it with offload. Day-zero ecosystem: Diffusers, ComfyUI, vLLM-Omni, SGLang, LightX2V; 14 fine-tunes and 18 quantizations within the first day. The core of this piece: every prior Qwen-Image shipped Apache 2.0 - 2.1 switches to the Qwen Research License (non-commercial only; commercial use needs a separate agreement; fine-tunes inherit the restriction and must display "Built with Qwen"). The community "license trap" pushback and the "License renders this model useless" thread are reported as-is. Qwen-Image-Bench 60.28 (seventh) is vendor-designed and vendor-run. Lineage section: 3.0 is the API-only fork, 2.1 is the downloadable line - the weights came back, the Apache license did not.

Oct 8, 202610 min read
Open Source

Univer Office SDK: Agents Edit Cells Without an AI SDK

Only what the dream-num/univer README, LICENSE and GitHub API fields support. Apache-2.0 covers the Web SDK core architecture and engine implementations; commercial plugins, the Server SDK and commercial AI SDK extensions are distributed under separate licences. Facade APIs such as FWorkbook, FWorksheet, FRange, getRange and setValues let an application or Agents Edit Cells directly with formula recalculation, while workbook.save() persists a JSON snapshot rather than XLSX import or export. No AI SDK is required anywhere in that path. Stars are cumulative GitHub API totals, not single-day growth.

Nov 16, 20269 min read
Open Source

llama.cpp and ggml Ship v1.0.0: Deprecated APIs Removed

Reads only what the GitHub API fields, the two v1.0.0 release notes, Discussion 1579, LICENSE and README of the ggml-org llama.cpp and ggml repositories can verify. Both v1.0.0 tags landed on November 13, 2026. It separates semantic versioning, the rule that a release tag requires the internal ggml copy to match the released one, the split between b-numbered nightlies and vX.Y.Z releases, and the 35 removed functions, RPC v9 and MIT notice-retention and AS IS terms. The August 2026 versioning discussion is not presented as a November announcement, and no figure from the unretrieved HF blog is quoted.

Nov 15, 20269 min read