Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableX: @MiniMax_AI→ original

MiniMax Releases Open-Weight H3-Image: 4K Image in 10 Seconds and Up to 16 References

MiniMax has released H3-Image, an image generation model that produces a 4K picture in about 10 seconds. Output resolution reaches 16 million pixels, and a single request can include up to 16 reference images, such as a product photo, a logo, and a style sample. The model can also edit existing pictures based on a text description, and its weights are open. For product teams, this is a chance to get print-quality output without a separate upscaler.

AI-processed from X: @MiniMax_AI; edited by Hamidun News
MiniMax Releases Open-Weight H3-Image: 4K Image in 10 Seconds and Up to 16 References
Source: X: @MiniMax_AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @MiniMax_AI ↗
Editor-in-Chief
Speeda full 4K image is generated in roughly 10 seconds
AccessAPI, MiniMax Code, and Hailuo AI, with weights published
Key metric #3Up to 16 references in one request, plus editing via text description

Image generation has long been stuck with two awkward trade-offs. Either the model is fast but outputs a mid-resolution frame that then has to be stretched with a separate upscaler, or the picture comes out sharp but takes minutes to render, and access is locked behind someone else's API. The Chinese company MiniMax says H3-Image solves both problems at once: 4K in 10 seconds and open weights.

What MiniMax Actually Claimed

The announcement appeared in the company's official X account and has drawn nearly 734 thousand views. The main points are these:

  • generating a 4K image takes about 10 seconds;
  • the maximum resolution is 16 million pixels (for comparison, a 4096×4096 square is close to this figure, while a standard 4K frame at 3840×2160 is about 8 million);
  • the model follows the text prompt closely and can edit existing images;
  • up to 16 reference images can be passed in a single request;
  • the model weights are open.

H3-Image is available through the MiniMax API, through MiniMax Code, and through the Hailuo AI service. The speed and resolution figures so far come from the company itself; the announcement includes no independent measurements.

Why 16 References Matter

The most practical detail here is not the resolution but the number of references. Imagine a clothing retailer that needs dozens of ad visuals. The request can include a photo of the jacket itself, the brand logo, a lighting sample from a previous shoot, a couple of photos of models, and a text description of the scene. The more reference images the model can hold at once, the less detail has to be painted in by hand, and the more consistent the style across a series.

The same applies to editing. If a model can not only create an image from scratch but also modify an existing one by text command (remove the background, change the color, add an object), it can be built into a single workflow instead of a chain of three or four tools.

Why Open Weights Matter

Open weights provide something closed services do not: you can deploy the model on your own servers, keep commercial materials from leaving your infrastructure, and manage generation costs yourself. For teams building marketplaces, photo editors, games, or ad builders, this is a serious argument, especially when the bill for hundreds of thousands of images a month grows faster than revenue.

There is a downside. The announcement does not say how large the model is, what GPUs are needed to achieve those 10 seconds, or under what terms commercial use is permitted. The "10 seconds for 4K" figure most likely refers to the company's own API running on its own hardware. On an ordinary workstation the timing will differ, and this has to be checked with a test rather than taken from a tweet.

What to Do in Practice

If you are building a product with image generation, it makes sense to spend a day assembling a small benchmark on your own tasks:

  1. Take 20–30 real requests, including difficult ones: with text in the image, multiple objects, and references.
  2. Run them through H3-Image and through the model you use now.
  3. Compare quality, response time, and the per-image API price, then estimate the cost of self-hosting.
  4. Check the license on the weights separately before building the model into a commercial product.

For the market, the news is also telling: Chinese labs keep releasing open multimodal models that, by their own claims, match closed Western services. Whether independent tests confirm this will become clear in the coming weeks, once developers start comparing H3-Image with familiar generators on identical prompts. For now, the industry has one more strong open option, and it is worth at least trying in real work.

Story

Key Questions Answered

What MiniMax Actually Claimed?

The announcement appeared in the company's official X account and has drawn nearly 734 thousand views. The main points are these: generating a 4K image takes about 10 seconds; the maximum resolution is 16 million pixels (for comparison, a 4096×4096 square is close to this figure, while a standard 4K frame at 3840×2160 is about 8 million); the model follows the text prompt closely and can edit existing images; up to 16 reference images can be passed in a single request; the model weights are open.

Why 16 References Matter?

The most practical detail here is not the resolution but the number of references. Imagine a clothing retailer that needs dozens of ad visuals. The request can include a photo of the jacket itself, the brand logo, a lighting sample from a previous shoot, a couple of photos of models, and a text description of the scene. The more reference images the model can hold at once, the less detail has to be painted in by hand, and the more consistent the style across a series. The same applies to editing.

Why Open Weights Matter?

Open weights provide something closed services do not: you can deploy the model on your own servers, keep commercial materials from leaving your infrastructure, and manage generation costs yourself. For teams building marketplaces, photo editors, games, or ad builders, this is a serious argument, especially when the bill for hundreds of thousands of images a month grows faster than revenue. There is a downside.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…