Understand this story
On November 10, 2026, MiniMax released the H3-based image model H3-Image and opened the model weights. The company said it can generate 4K×4K, or 16-megapixel, images within 10-20 seconds, supports strong instruction following, image editing, and up to 16 reference images, and is now available through the API, MCode, and Hailuo AI. On November 11, Moore Threads announced that the MTT S5000 and the MUSA software stack completed Day-0 adaptation on the day of open sourcing; on the same day, MiniMax officially showcased generation results from multiple sets of prompts, including converting complex information into charts and evidence boards. On November 12, MiniMax Design said the model's handling of AAA game worlds and 3D environments exceeded expectations. On November 13, the Arena leaderboard showed that H3-Image ranked first among open-source models with 1,365 points in the multi-image editing arena and 11th overall. On November 14, Artificial Analysis said it became the top-ranked open-weight image model on AA-Image-Editing v2.1 and AA-Image-T2I v2.1; on the same day, PixVerse announced the launch of MiniMax H3 Image.
AI-generated from public reports
Latest update · Discussion
Ayi AI Notes joined this discussionDevelopments
9 confirmed developments
Artificial Analysis says MiniMax H3-Image tops the open-weight image model leaderboard
Source:Artificial Analysis- Source:PixVerse
Arena calls MiniMax H3-Image the top open-source model in the multi-image editing arena
Source:ArenaMiniMax Design says H3-Image's performance on 3D scenes exceeds expectations
Source:MiniMax Design (H3)Moore Threads completes adaptation for MiniMax H3-Image model
Source:IT Home · AI- Source:MiniMax
MiniMax demonstrates H3-Image converting complex information into charts and posters
Source:MiniMax- Source:MiniMax
- Source:MiniMax
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 15
Ayi AI NotesRelated discussion
Ayi AI Notes joined this discussion
This post contributes to heat; it is not a formal report or confirmed fact.
Recode China AIFirst-party
Jensen Huang steps down as NVIDIA CEO, Alibaba's chip subsidiary files for Hong Kong listing, Qwen4-Max makes its debut
NVIDIA announced that Jensen Huang will step down as CEO and from the board of directors on November 11, with CFO Colette Kress serving as interim CEO; the board has launched an independent investigation into allegations concerning Huang himself and NVIDIA's China business. Alibaba disclosed on November 10 that its chip subsidiary T-Head confidentially submitted an application for a Hong Kong listing on September 11, with offering terms not yet determined. Alibaba also previewed Qwen4-Max, the first model in the next-generation Qwen4 family, on the same day; it supports text, image, and video input and a 1 million token context window.
AYi AI Notes
After MiniMax-H3-Image was open-sourced, the community quantized it into 27 versions, with Q2_K compressed to 6.4G
MiniMax's open-source 33.1B image generation model MiniMax-H3-Image has an original size of 68.65 GB; the community has already produced 27 quantized versions, with over 110,000 downloads. The author compiled actual test feedback stating that the Q4 version is compressed to 10.78G, Q2_K to only 6.4G, and Unsloth's FP8 version is 18.8G with image quality close to the original. The model supports native 4096×4096 direct output, 6-step sampling without enabling CFG, and can take 16 reference images at once to maintain character consistency; ComfyUI's GGUF nodes and Mac's MLX weights are already ready.
Nov 14
PixVerse
MiniMax H3 Image Launches on PixVerse
PixVerse has announced that MiniMax H3 Image is now live, supporting up to 16 reference images, realistic visuals at up to 4K, and complex prompts and detail editing. For the first 7 days after launch, all users enjoy 50% off, and Ultra members enjoy 70% off.
TMTPost · AGI
Hands-on Test of MiniMax H3-Image, GPT Image 2.6 and Seedream 5.5: Who Is Faster and Who Is More Accurate
The author tested three new image models—MiniMax H3-Image, OpenAI GPT Image 2.6 and ByteDance Seedream 5.5—via Hailuo AI, ChatGPT and Doubao. MiniMax generated images in 10-15 seconds in Hailuo AI's high-speed mode, leading on speed, but under complex prompts it showed clipping and unfused faces; OpenAI ranked first on both the Artificial Analysis text-to-image and image-editing leaderboards, with the most accurate understanding and modification of complex instructions; Seedream 5.5 missed changing one face during local editing, and a second revision took about 7 minutes. The article also analyzes the three companies' strategies: OpenAI relies on the ChatGPT entry point, ByteDance relies on content scenarios such as Jimeng, and MiniMax seeks cost-sensitive customers with speed, low prices and open weights.
Arena
Arena Weekly: Claude Fable 5.2 Tops Agent Arena, Gemini 4…
This week's Arena leaderboard shows that Anthropic's Claude Fable 5.2 (Max) tops Agent Arena, achieving a +30.5% net improvement based on 7.8k real sessions. Google DeepMind's Gemini 4 Hydrogen (High) is the fastest model in the Agent Arena top ten, with a typical single-step latency of 7.3 seconds. Meta's smallest model, Muse Ember (Max), entered Code Arena: WebDev for the first time, ranking 37th with 1609 points and a median cost of $0.03 per task; thinkymachines' Inkling 2 Small, after opening its weights, became the top US open-source model on that leaderboard with 1647 points; MiniMax H3-Image ranked first among open-source models and 11th overall on Multi-Image Edit with 1365 points; Baidu's ERNIE 5.5 Flash Preview, with 1668 points (AutoEval), is closing in on second place among open-source models on that leaderboard, with a 1M token context window.
Artificial Analysis
MiniMax H3-Image Tops AA Open-Weights Rankings for Image Editing and Text-to-Image
Artificial Analysis's evaluation shows that MiniMax H3-Image has become the top-ranked open-weights image model on both the AA-Image-Editing v2.1 and AA-Image-T2I v2.1 leaderboards. In the overall rankings, H3-Image places 13th in image editing and 25th in text-to-image; the accelerated H3-Image Highspeed version, available only through the MiniMax API, places 14th and 24th respectively.
Nov 13
36Kr · AI
Hands-on comparison of MiniMax H3-Image, GPT Image 2.6, and Seedream 5.5
The author used the same set of complex prompts to test MiniMax H3-Image, GPT Image 2.6, and Seedream 5.5 on Hailuo AI, ChatGPT, and Doubao. H3-Image's high-speed mode generates images in about 10-15 seconds, faster than the roughly 1 minute of the other two, but under complex instructions it showed issues such as clipping and failure to blend the face with the tanghulu; GPT Image 2.6 had the most accurate comprehension and modification, while Seedream 5.5 missed changes in local editing and took about 7 minutes. The article also mentions that MiniMax opened the H3-Image weights on the evening of November 10, priced at 50%-70% of mainstream tools in the same tier.
ArenaSelected
MiniMax H3-Image Tops the Multi-Image Editing Arena as the No. 1 Open-Source Model with a Total Score of 1365
The Arena leaderboard shows that MiniMax H3-Image ranks first among open-source models and 11th overall in the multi-image editing arena with 1365 points, tied with Nano Banana 2 and Nano Banana Pro, and is the only open-source model in the top 20. In single-image editing, it ranks second among open-source models with 1359 points, tied with Qwen-Image-2.1; in text-to-image, it ranks third among open-source models with 1197 points, tied with Ideogram 4.0. According to MiniMax Labs, the model is the fastest 2K image editing model, taking 10-15 seconds to generate a 2K edit, 3 times faster than Nano Banana 2.1.
Nov 12
QbitAISelected
MiniMax Open-Sources Image Model H3-Image, Supporting Native 4K Generation
MiniMax has open-sourced its image model H3-Image, which supports native 4K generation at 16MP (4096×4096), with image output on the web taking roughly 10-20 seconds in the author's experience. In hands-on testing, the author found it approaches real photographs in details such as facial pores and fine hairs, and specular water highlights; it can also derive orthographic three-view drawings and mechanical exploded views from product appearance images, and correctly render complex integral expressions. API pricing is 50% to 70% of mainstream competitors, with an additional 50% off 16MP requests in the first week.
36Kr · AISelected
MiniMax Releases Open-Source Image Model H3-Image, Generating 4K Images in 10 Seconds with Open Weights
MiniMax has released its next-generation image generation model H3-Image and simultaneously opened the model weights. The high-speed version can generate a 4096×4096-pixel (16MP) 4K image in about 10 seconds, roughly 1/5 the time of mainstream models on the market. The model supports fusion of up to 16 reference images and can handle complex instructions and structured information. The author tested it via MiniMax Hailuo and MiniMax Code, covering tasks such as realistic photography, style transfer, Chinese poster layout, mechanical exploded views, and industrial three-view drawings. Among these, the three-view drawings showed deviations in the display screen width ratio, which were later located and corrected by MiniMax Code calling a script.
MiniMax Design (H3)First-party
MiniMax responds to H3-Image's effect in generating 3D game scenes
MiniMax Design reposted a creator's work using MiniMax-H3-Image to generate AAA-style 3D game environments, saying it didn't expect the model to handle it so well, and praising the creator community for continuing to bring surprises.
TMTPost · AGI
MiniMax H3 Image vs. GPT Image 2.6 hands-on comparison: cheaper and faster, but still behind on complex instructions
The author compares MiniMax H3 Image and GPT Image 2.6 across six rounds of hands-on testing. H3 Image has advantages in image realism, generation speed and price—a 4K 16MP image costs about 0.2261 yuan per image with the first-week discount—but it still falls short of GPT Image 2.6 in executing complex instructions, precise quantity control and commercial poster design. H3 Image is derived from the H3 multimodal model, supports text-to-image, image editing and up to 16 reference image inputs, and its model weights have also been released. The author also notes that MiniMax's image capabilities are scattered across different entry points such as Hailuo and MiniMax Code, so the experience for ordinary creators is less coherent than the ChatGPT ecosystem.
Nov 11
MiniMaxFirst-party
MiniMax H3-Image demonstrates generating 16 reference images in a single pass
MiniMax officially used MiniMax H3-Image to demonstrate multiple types of image generation: a single generation can accommodate 16 reference images, suitable for material arrangement, spatial composition, and product packaging; the model can also turn complex information into charts and posters, for example integrating a case narrative into a single evidence board with a timeline, annotations, and connecting lines. The official team also released several complete prompts, including a Dunhuang mural-style space station, slow-shutter tracking shots, and dreamcore style.
MiniMaxFirst-party
MiniMax H3-Image Turns Complex Information into Charts and Evidence Boards
MiniMax officially demonstrated H3-Image's image generation capabilities: it can turn complex information into charts and posters, and can also consolidate a long case narrative into a single evidence board, grouping, labeling, and connecting timelines, positions, and testimonies. The official example required the model to solve for the only true culprit among six suspects and generate a complete 16:9 cinematic noir detective-style reasoning board. The same thread also mentioned that a single generation can reference 16 images, suitable for asset arrangement, spatial composition, and product packaging.
MiniMaxFirst-party
MiniMax H3-Image Demonstrates Multiple Types of Image Generation Capabilities
MiniMax officially used MiniMax H3-Image to demonstrate a variety of image generation scenarios, including dreamcore-style photography, slow-shutter panning sports photos, detective reasoning evidence boards, and a Dunhuang mural-style space station maintenance image. The model can reference up to 16 images in a single generation, making it suitable for tasks such as material arrangement, spatial composition, and product packaging. The official team also released the complete prompts for each example and invited users to share their own works and prompts.
MiniMaxFirst-party
MiniMax Showcases H3-Image Image Generation Capabilities and Prompt Examples
MiniMax officially released a series of examples showcasing the image generation results of MiniMax H3-Image, along with complete prompts. The examples cover motion blur from slow-shutter panning shots, dream-style surreal photography, an infographic that turns a detective reasoning puzzle into an evidence board, and a space station scene in Dunhuang mural style. The official announcement also noted that a single generation can reference 16 images, making it suitable for scenarios such as material arrangement, spatial composition, and product packaging.
MiniMaxFirst-party
MiniMax H3-Image Showcases Results from Multiple Sets of Prompts
MiniMax officially used MiniMax H3-Image to showcase the results of image generation from multiple sets of prompts, covering themes such as slow-shutter panning sports photography, Dreamcore surreal scenes, detective reasoning evidence boards, and Dunhuang mural-style space stations. The model can reference up to 16 images in a single generation, making it suitable for scenarios such as material arrangement, spatial composition, and product packaging.
AYi AI Notes
Claude Fable 5.2 prompt hands-on test: MiniMax H3-Image image generation
The author used nine portrait prompt rules summarized by Claude Fable 5.2 to run a hands-on test of H3-Image, released by MiniMax the same day, generating a nine-grid set of at-home portraits, and reported that details such as strands of hair, the faint sheen on the collarbone, and the ribbing of cotton fabric were clear, with the face remaining largely consistent across different poses. H3-Image is derived from H3 and supports text-to-image, image reference, and image editing; it can generate 4K images and produces an image in about 10 seconds. The author speculates that its approach is to distill the video foundation model after removing audio, thereby transferring the video model's understanding of three-dimensional lighting to static portraits.
Testing Catalog
AI Daily Brief: Gemini 4 Carbon launches, Anthropic kicks off IPO roadshow
Gemini 4 Carbon is now available to Ultra subscribers in the Gemini app, Antigravity, and AI Mode, with API pricing at $2/$10 per 1M tokens, and it scored 60.4% on Cursor's CursorBench high-compute tier. Anthropic has kicked off its IPO roadshow, with a price range of $70-$78 per share; Stripe has completed its acquisition of OpenRouter. MiniMax H3-Image has been released with open weights, generating 4K images in about 10 seconds.
English text is being prepared.
MiniMax Open-Sources H3-Image Image Generation Model, Generating 4096×4096 Images in 10 Seconds
MiniMax has released and opened the weights of the H3-Image model, which is derived from H3 and supports text-to-image generation, image reference, and image editing. The company claims it can generate a 4096×4096-pixel, approximately 16MP image within 10 seconds, at a speed up to 5 times that of mainstream offerings on the market, and supports up to 16 reference images. During the first week of release, generation requests at 16MP and above are eligible for a 50% discount on official pricing.
IT Home · Artificial Intelligence
Moore Threads MTT S5000 Completes Day-0 Adaptation for MiniMax H3-Image Image Generation Model
Moore Threads announced that its MTT S5000 AI training-and-inference integrated intelligent computing card and MUSA software stack completed adaptation on the same day MiniMax H3-Image was open-sourced. H3-Image is derived from H3 and supports text-to-image, image reference, and image editing. It can generate a high-resolution image of 4096 × 4096 pixels, approximately 16MP, in under 10 seconds, at a speed up to 5 times that of mainstream levels on the market. Moore Threads performed full-link adaptation and deep optimization based on the MTT S5000, reusing components such as SGLang, MATE, and MCCL.
Nov 10
MiniMax Design (H3)First-party
MiniMax Releases Open-Source Weight Image Model H3
MiniMax Design (H3) has released an open-source weight image model aimed at the community, and provides four usage entry points: Hugging Face, API, Hailuo AI, and MiniMax Code. The official description calls the model "one idea, many uses" and invites users to create.
MiniMax Design (H3)First-party
MiniMax Releases Open-Weight Image Model H3-Image
MiniMax has released the image model H3-Image, supporting 16-megapixel resolution, up to 16 reference images, and generating 4K images in 10 seconds, released as open weights. The model is now available via API, MiniMax Code, and Hailuo AI, with 50% off generations above 16MP during launch week, and the weights are hosted on Hugging Face.
MiniMaxFirst-party
MiniMax Releases H3-Image Image Model, Generating 4K Images in 10 Seconds with Open Weights
MiniMax has released the image model H3-Image, which can generate 4K images within 10 seconds at a resolution of 16 million pixels, supports precise prompt following and editing, and can reference up to 16 images. The model has open weights and is now available via the MiniMax API, MiniMax Code, and Hailuo AI; during the release week, generations above 16MP enjoy a 50% discount.
MiniMaxFirst-partySelected
MiniMax Releases Image Model H3-Image and Opens Model Weights
MiniMax has released H3-Image, an image model based on H3, and opened its model weights. According to the official announcement, it can generate 4K × 4K, i.e., 16-megapixel images in 10-20 seconds, and supports strong instruction following, image editing, and up to 16 reference images. The model is now available via API, MCode, and Hailuo AI, and generations above 16 megapixels enjoy a 50% discount during the release week.
Heat trend for this story
- Current
- 9
- Period peak
- 71 Nov 12 21:00
- 24-hour change
- -36%
Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.