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: @arena→ original

Open-Weight MiniMax H3-Image Ties Nano Banana 2 in Photo Editing With 1365 Arena Points

MiniMax has released H3-Image, an image editing and generation model whose weights can be downloaded and run locally. On Arena, where people blindly pick the better of two images, it scored 1365 points in multi-image editing: 11th place and a tie with Nano Banana 2 and Nano Banana Pro. It is the only open model in the top 20. That means product photos, backgrounds and style references can be combined on your own servers, without a paid API.

AI-processed from X: @arena; edited by Hamidun News
Open-Weight MiniMax H3-Image Ties Nano Banana 2 in Photo Editing With 1365 Arena Points
Source: X: @arena. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @arena ↗
Editor-in-Chief
Single-image editing1359 points, second place among open models
Text-to-image generation1197 points, third place among open models
Key metric #32K editing in 10–15 seconds, three times faster than Nano Banana 2.1 (according to MiniMax)

Imagine an online store that needs a product card: it has a photo of the item, a background, a style reference and a shot of a model, and needs one clean image as the output. Until now, only closed models reliably handled this kind of multi-photo editing, and you had to pay for someone else's API. Now MiniMax has released H3-Image with open weights, and on the Arena leaderboard it scored 1365 points specifically in multi-image editing. That puts it 11th among all models, open and closed, and tied with Nano Banana 2 and Nano Banana Pro. It is the only open model in the top 20.

What the points mean

Arena is a platform where users blindly compare the outputs of two models and vote for the better one. The votes produce ratings similar to a chess rating: the higher the score, the more often a model beats its rivals. A tie means the difference falls within statistical error. That does not mean the images are identical, but according to live human judgments, an open model is sitting alongside the paid leaders.

In the Multi-Image Edit category, where the result must be assembled from several source images, H3-Image scored 1365 points. In the other categories the picture is more modest, but still strong:

  • single-image editing: 1359 points, second place among open models, on par with Qwen-Image-2.1;
  • text-to-image generation: 1197 points, third place among open models, on par with Ideogram 4.0.

Arena itself notes that open weights are starting to narrow the gap with leading closed models.

Speed and resolution

MiniMax puts particular emphasis on speed. According to the lab, H3-Image is the fastest editing model at 2K resolution. Such an edit takes 10–15 seconds, which is three times faster than Nano Banana 2.1. These are MiniMax's own figures; there is no independent comparison in the material, so the number should be checked on your own tasks.

A few more characteristics from the announcement:

  • 4K generation (about 16 million pixels) in roughly 10 seconds;
  • up to 16 reference images in a single request;
  • precise prompt adherence and realism, which the authors describe as the main strengths.

Sixteen references are not only for designers. This is a scenario for marketplaces, advertising banners and social media, where the same product or character must look consistent across different frames.

What it means for developers

The model is already available through the MiniMax API, the MiniMax Code service and the Hailuo AI app. But the key news is the open weights. They can be downloaded, deployed on your own servers, fine-tuned to your own style, and user photos never have to be sent to an external service. For companies with strict data privacy requirements, that is a serious argument.

There are caveats too. The original announcement does not specify the license, the model size or the VRAM requirements. Without these, it is impossible to estimate what running it yourself will cost or whether the model can be used in a commercial product. The 10–15 second speed was most likely measured on powerful server hardware, and on ordinary GPUs the result will be different.

For product teams, the takeaway is simple. If you are building image editing features, add H3-Image to the shortlist alongside the closed models and run it on your own examples. General-purpose rankings may not match quality on photos from your specific domain. If the results turn out comparable, open weights give you more control over cost and data.

The gap between open and closed image models is shrinking: in multi-photo editing, an open model is already in the same tier as the leaders. The remaining question is whether this result holds up in real products.

Story

Key Questions Answered

What the points mean?

Arena is a platform where users blindly compare the outputs of two models and vote for the better one. The votes produce ratings similar to a chess rating: the higher the score, the more often a model beats its rivals. A tie means the difference falls within statistical error. That does not mean the images are identical, but according to live human judgments, an open model is sitting alongside the paid leaders.

What are the key details on Speed and resolution?

MiniMax puts particular emphasis on speed. According to the lab, H3-Image is the fastest editing model at 2K resolution. Such an edit takes 10–15 seconds, which is three times faster than Nano Banana 2.1. These are MiniMax's own figures; there is no independent comparison in the material, so the number should be checked on your own tasks.

What it means for developers?

The model is already available through the MiniMax API, the MiniMax Code service and the Hailuo AI app. But the key news is the open weights. They can be downloaded, deployed on your own servers, fine-tuned to your own style, and user photos never have to be sent to an external service. For companies with strict data privacy requirements, that is a serious argument. There are caveats too. The original announcement does not specify the license, the model size or the VRAM requirements.

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…