MiniMax H3 Image ships with open weights and an API on launch day; 4K output reported at 10–20 seconds
According to a Xinzhiyuan report dated November 10, 2026, MiniMax released H3 Image, an image model, with open weights and a concurrently open API on the same day. The report says it directly outputs 4096×4096 images, accepts up to 16 reference images, and supports prompts of up to 7,500 characters. The report does not give the weight license, parameter count, or deployment requirements, and this version is not in the current KimiSeek catalog.
What changed
According to a Xinzhiyuan report dated November 10, 2026, MiniMax released H3 Image, an image-generation model, opening the weights on launch day and concurrently opening an API, with integration into MiniMax's own MiniMax Code, MiniMax open platform, and Hailuo AI. The specifications listed in the report are: direct generation of 4096×4096 images, acceptance of up to 16 reference images, and a prompt limit of 7,500 characters. The official positioning given is SOTA among open-source image models, with overall quality in the global first tier; the report says this version emphasizes open source, high resolution, and speed, and singles out instruction-following as a point worth watching. On speed, the report says a 4K image takes about 10 to 20 seconds, and compares this with mainstream models that mostly work at 1–4 megapixels, take minutes, and at fastest take around 50 seconds. The report also relays trial observations from several bloggers: generating a full set of 4K character design sheets from one photo, redrawing old game screenshots into a AAA look, and relatively stable layout for large-text posters with occasional errors on small text; it also mentions that a blogger had the model generate CAD-style images and mechanical exploded views, but gives no clear assessment of the exploded views. On sign-language gestures, the report relays a third-party statistic claiming H3 Image scored 48.1% accuracy versus 20.0% for a comparison model, Moss. The report also mentions that MiniMax open-sourced a video model, H3, three months earlier, and that Nvidia's research team, Pika, and fal had retrained derivative models based on it.
What it means for developers
For integrators, this report changes at least three inputs to decision-making. The following is conditional analysis based on the report's content, not a verified conclusion. First is output resolution: native 4096×4096 creates an alternative to the existing pipeline of low-resolution generation followed by upscaling and stitching, but whether the alternative pays off depends on whether the deliverable really needs that pixel count; in web or feed scenarios, the extra mainly amounts to generation time and storage overhead. Second is response time: the report places 4K generation in the 10–20 second range and says mainstream models are mostly at the minute level or around 50 seconds. If that holds under real load, the waiting experience for interactive editing would be clearly different; but the report gives no information on concurrency, queuing, or failure rates, and per-image time cannot be directly converted into capacity planning. Third is input method: 16 reference images and a 7,500-character prompt mean cross-image consistency can be delegated to multi-image conditioning instead of pre-compositing locally, at the cost of larger request bodies and prompt maintenance shifting to the product side. Open weights and an open API appearing on the same day superficially offer two deployment paths, but the report does not cover the license, weight format, VRAM usage, or inference framework; what can be confirmed for now is only that a calling path exists, and the feasibility of local self-hosting remains unknown. The reported performance on in-image text, sign-language gestures, and CAD-style images comes from blogger trials and relayed third-party statistics, and is a capability claim to be self-tested, not a reproducible evaluation conclusion.
Integration and evaluation
The following are conditional recommendations and evaluation steps, judgments based on the report, not report facts. Rule one, whether to drop the upscaling stage: the corresponding reported change is native 4096×4096 output; the choice is whether to keep low-resolution generation plus upscaling or switch to native 4K; the precondition is whether the deliverable needs detail at 100% zoom, such as print, large crops, or e-commerce hero-image details; the measurement is to run the same batch of 30 prompts through both pipelines, spot-check text edges and geometric straight lines at 100% zoom, record the proportion needing rework and the rate of upscaling seams, and switch only when the native path's rework rate is clearly lower. Rule two, synchronous waiting or an async queue: the corresponding reported change is the claim of about 10–20 seconds for 4K generation; the choice is whether to put generation inside the user request or use a queue plus callback; the precondition is target concurrency, acceptable wait time, and vendor-side concurrency and timeout limits, which the report does not provide; the measurement is to test end-to-end p95 time (including queuing) and timeout failure rate at target concurrency, then compare with your service agreement. Rule three, how to use the multi-reference channel: the corresponding reported change is 16 reference images and a 7,500-character prompt; the choice is whether to keep compositing multiple assets into one input image or switch to multi-image conditioning; the precondition is whether the task requires cross-image consistency, such as multi-view character sheets or multi-angle product displays; the measurement is to fix one character design and generate the same number of different compositions with 1 and 8 references respectively, use your own consistency check to count pass rates, and record per-request time and retry rate. Rule four, whether to put it in the candidate pool: the corresponding reported change is the report's description of instruction-following and blogger tests; the choice is whether to hand strongly constrained tasks such as in-image text and CAD-style images to it; the precondition is whether the task contains strings that must be exactly correct, such as poster copy or parameter labels; the measurement is to prepare 20–30 prompts with fixed strings and layout constraints, do character-level comparison, and count the exact-match rate. The relayed 48.1% and 20.0% figures have unknown methodology and cannot be used as a baseline.
Limits and open questions
The report does not give H3 Image's weight license, parameter count, weight format, VRAM usage, or inference framework, nor API pricing or quotas; these should be confirmed through MiniMax's official channels, and pricing and billing should be based on the official pricing page. It should be clear that H3 Image is not currently in KimiSeek's live model catalog; the open API mentioned above refers to the vendor's channel and does not mean this site provides the model. Open weights on launch day also do not equal an open-source license: weight availability and whether commercial use or redistribution is allowed are two different things, and the report only covers the former. The speed comparison (10–20 seconds versus about 50 seconds) and the sign-language figures of 48.1% and 20.0% all come from relayed claims in the report; methodology, sample size, and hardware environment are not given and cannot be independently reproduced. In addition, H3 Image is an image model and is not the same model as the previously open-sourced video model H3 of the same name; parameters, licenses, and capabilities cannot be cross-referenced. The other products and timelines mentioned at the start of the report are narrative background for that report and are only relayed here, not verified.