# MiniMax H3 > MiniMax H3 is MiniMax's general-purpose multimodal video model, announced on 31 July 2026 as the successor to Hailuo 01 and 02, that makes clips of up to 15 seconds with native stereo sound from text, a start image, or up to 9 images, 3 videos and 3 audio clips as references. Its base weights are on Hugging Face as MiniMaxAI/MiniMax-H3 (33B parameters). On Kenerate AI it runs as MiniMax H3 (480p or 768p, from 10 credits a second) and MiniMax H3 Pro (768p or 2K, 30 or 42 credits a second), so a 5-second 768p Pro clip costs 150 credits. URL: https://kenerateai.com/models/minimax/minimax-h3 Publisher: Kenerate AI (https://kenerateai.com) Last updated: 2 November 2026 ## Key facts - Official name: MiniMax H3 - Hugging Face: MiniMaxAI/MiniMax-H3 - Maker: MiniMax (Hailuo AI) - Announced: 31 July 2026 - Size: 33B dense transformer - Clip length: 3–15 s (H3) · 4–15 s (H3 Pro) - Resolution: 480p · 768p · 2K - References: 9 images · 3 videos · 3 audio - Sound: Native stereo - Where to run it: Hailuo AI app · MiniMax API · Kenerate - Billing: Per second, no subscription ## Steps 1. Open the video studio — Go to /app/video and pick MiniMax H3 Pro (the button on this page selects it) or MiniMax H3 for cheaper drafts. Video needs any one-time credit pack. 2. Add a start image or references (optional) — Upload a start frame — plus an end frame on H3 Pro — or switch to Reference to Video and add up to 9 images, 3 clips and 3 audio files. Leave everything empty for text to video. 3. Describe the shot — Write subject, action, camera move and sound. Put spoken lines in quotes and say who speaks; refer to references by their order. 4. Pick length, resolution and ratio — Choose 3–15 seconds (4–15 on Pro), 480p, 768p or 2K, and an aspect ratio. The credit cost updates before you generate. ## In depth ### What is MiniMax H3, in plain terms? MiniMax H3 is one model that understands text, images, video and audio together and outputs a video with its soundtrack in the same pass [1]. MiniMax positions it for advertising, branding, e-commerce, product design, UI/UX and games, and highlights instruction following, accurate text and brand rendering, multi-shot text-to-video and video-to-video motion transfer [1]. On price, MiniMax claims its 2K per-second rate is less than a third of mainstream models', and its 768p rate less than half their 720p rate [1]. ### What's new in MiniMax H3 compared with Hailuo 02? MiniMax's announcement lists four technical changes behind H3 [1]: - Contextual Omni Representation — captions that compress source material needing about 100K tokens down to roughly 4K on average. - H3-VAE — a reworked tokenizer with a high compression ratio and a 4x gain in effective sequence length, which MiniMax credits for native 2K. - H3-Omni Transformer — nearly 30% higher training throughput. - In-Context Regeneration — the base model re-renders its own low-resolution draft at 2K instead of handing it to a separate super-resolution model. ### MiniMax H3 vs H3 Pro vs H3 Max Three names circulate for H3. Only the first two run on Kenerate: - MiniMax H3 (kenerate/minimax-h3) — 3–15 s, 480p or 768p, 7 aspect ratios, text, image and reference to video, 10–20 credits a second. - MiniMax H3 Pro (kenerate/minimax-h3-pro) — MiniMax's own minimax/h3 API, 4–15 s, 768p or 2K, start and end frame, 30–42 credits a second. - H3 Max — per Morphic, a fal post-trained version of the open weights at 480p or 768p without references; MiniMax hasn't announced it [5]. Not offered on Kenerate. ### Which MiniMax H3 tier to use Both tiers run the same model family and take the same references [1][2]; they differ in resolution, length range and price. Prices below are per 5-second clip on Kenerate. - MiniMax H3 at 480p — 50 credits. Testing a prompt, social drafts, ultra-wide 21:9 or tall 9:21 framing. - MiniMax H3 at 768p — 100 credits. Finished social clips on a budget. - MiniMax H3 Pro at 768p — 150 credits. The everyday default, with start and end frames. - MiniMax H3 Pro at 2K — 210 credits. Final renders where detail matters. ### Prompting MiniMax H3 with references H3 reads every input in one context [1], so say what each reference is for: "the woman in image 1 wears the jacket from image 2 and dances like the clip". Keep reference clips short — on Kenerate their seconds are billed like output seconds — and add audio only alongside an image or a clip. Put spoken lines in quotes and name the language; ComfyUI lists 11 dialogue languages [2]. ### How good is MiniMax H3? The Batch, citing Artificial Analysis's human-preference leaderboards in August 2026, reported H3 first for video editing and in the top three for text-to-video and image-to-video alongside Gemini Omni Flash and Seedance 2.0 [3]. Leaderboards move every few weeks, so test your own prompt with a 5-second 480p MiniMax H3 draft (50 credits) before paying for a 2K Pro render. ### Self-hosting vs running it on Kenerate The base weights are downloadable, but the model is 33B parameters with a 32B text encoder, the licence asks users in the US, UK, EU and South Korea to apply first, and the 2K regenerate module isn't in the open release [2][3]. Running H3 on Kenerate needs no GPU or licence application, and sits on the same credit balance as Seedance, Kling, Veo and Wan. References: [1] MiniMax H3 announcement — MiniMax: https://www.minimax.io/blog/minimax-h3 [2] MiniMax H3 model notes — ComfyUI Wiki: https://comfyui-wiki.com/en/models/minimax/minimax-h3 [3] MiniMax's state-of-the-art video model is only minimally open — The Batch: https://www.deeplearning.ai/the-batch/minimaxs-state-of-the-art-video-model-is-only-minimally-open [4] MiniMax H3 day-0 support in ComfyUI — Comfy blog: https://blog.comfy.org/minimax-h3-day-0-support-in-comfyui/ [5] MiniMax H3 Max model notes — Morphic (as of 2 Nov 2026): https://morphic.com/resources/models/minimax-h3-max ## FAQ Q: What is MiniMax H3? A: MiniMax H3 is MiniMax's general-purpose multimodal video model, announced on 31 July 2026 as the successor to its Hailuo 01 and 02 models. It reads text, images, video and audio in one context and generates clips of up to 15 seconds with native stereo sound, at up to 2K. Its base weights are on Hugging Face as MiniMaxAI/MiniMax-H3. Q: Is MiniMax H3 the same as Hailuo 3 or MiniMax M3? A: MiniMax's announcement calls it MiniMax H3 and describes it as the model after Hailuo 01 and 02; it never names it Hailuo 3, though some sites do. Hailuo AI is MiniMax's consumer video app. MiniMax M3 is a different line: the same post says MiniMax plans to bring M-series capabilities into the H series, so H3 is not M3. Q: What is MiniMax H3 Max? A: As of 2 November 2026, H3 Max is not a model MiniMax has announced. Morphic's model page describes it as fal's post-trained version of H3's open weights: text or image to video at 480p or 768p, 5–15 seconds, with no 2K stage and no reference inputs. Kenerate doesn't offer it; for 2K and references use MiniMax H3 Pro. Q: What's the difference between MiniMax H3 and H3 Pro on Kenerate? A: MiniMax H3 Pro is MiniMax's own minimax/h3 API: 4–15 second clips at 768p or 2K, with a start and end frame, for 30 or 42 credits a second. MiniMax H3 is a cheaper hosted version at 480p or 768p (10 or 20 credits a second), 3–15 seconds and seven aspect ratios including 21:9 and 9:21. Both take the same 9 image, 3 video and 3 audio references. Q: How much does MiniMax H3 cost on Kenerate? A: It's billed per second. MiniMax H3 costs 10 credits a second at 480p and 20 at 768p, so a 5-second clip is 50 or 100 credits; in reference mode it's 12.5 or 31.25 a second plus 5 per reference image. H3 Pro costs 30 a second at 768p and 42 at 2K — 150 or 210 credits for 5 seconds. Credits come in one-time packs from $15 for 1,300 credits. Q: Is MiniMax H3 free? A: Not on Kenerate's free plan — video generation needs a paid plan, which means buying any one-time credit pack, with no subscription. The base weights are free to download from Hugging Face under the MiniMax H3 Community License, but you need a large GPU to run a 33B model and some regions must apply to MiniMax first. Q: Is MiniMax H3 open source? A: Partly. MiniMax published the base weights on 3 August 2026 under its own MiniMax H3 Community License, not a standard open-source licence. The Batch reports that the licence excludes the US, UK, EU and South Korea unless users apply, bars distilling other models from H3's output, and that modules such as the H3-Regenerate-2K upscaler stay closed. Q: Where can I use MiniMax H3? A: MiniMax launched H3 in its Hailuo AI app and on its own API platform; several API resellers list it too, and the base weights can be self-hosted. On Kenerate you run it in the browser from /app/video as MiniMax H3 or MiniMax H3 Pro, on the same credit balance as Seedance, Kling, Veo and Wan, with no MiniMax account or API key. Q: How long can MiniMax H3 videos be, and what can I feed it? A: Up to 15 seconds per clip: 3–15 seconds on MiniMax H3 and 4–15 seconds on H3 Pro, in whole seconds. You can start from text alone, a start image (plus an end frame on H3 Pro), or reference mode with up to 9 images, 3 video clips and 3 audio clips — up to 15 seconds of reference video and 15 of reference audio in total, and audio only alongside an image or clip. Q: Does MiniMax H3 make sound? A: Yes. MiniMax says H3 models voice, sound effects and music jointly and outputs native stereo audio; ComfyUI's notes list 32 kHz stereo and dialogue in 11 languages. The clips on this page are MiniMax H3 Pro image-to-video renders with the model's own sound. ## Related pages - Hailuo AI video generator: https://kenerateai.com/hailuo-ai-video-generator - MiniMax Image: https://kenerateai.com/model/minimax - Seedance 2.0: https://kenerateai.com/models/bytedance/seedance-2-0 - Kling 3.0: https://kenerateai.com/models/kwaivgi/kling-v3-0 - MiniMax H3 prompt guide: https://kenerateai.com/blogs/minimax-h3-prompt-guide - All AI video models: https://kenerateai.com/ai-video-models ## Sources - MiniMax H3 announcement — MiniMax: https://www.minimax.io/blog/minimax-h3 - MiniMax H3 model notes — ComfyUI Wiki: https://comfyui-wiki.com/en/models/minimax/minimax-h3 - MiniMax's state-of-the-art video model is only minimally open — The Batch: https://www.deeplearning.ai/the-batch/minimaxs-state-of-the-art-video-model-is-only-minimally-open - MiniMax H3 day-0 support in ComfyUI — Comfy blog: https://blog.comfy.org/minimax-h3-day-0-support-in-comfyui/ - MiniMax H3 Max model notes — Morphic (as of 2 Nov 2026): https://morphic.com/resources/models/minimax-h3-max