--- title: MiniMax H3 is open-weight, 2K video, and 70% cheaper... description: MiniMax released H3, an open-weight video generation model priced at $0.081 per 15-second 2K clip — significantly undercutting Sora and Veo. H3 accepts up to... canonical: https://daily.dev/posts/minimax-h3-is-open-weight-2k-video-and-70-cheaper-than-the-competition-ppgcxunyx twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: MiniMax H3 is open-weight, 2K video, and 70% cheaper than the competition | daily.dev og:description: MiniMax released H3, an open-weight video generation model priced at $0.081 per 15-second 2K clip — significantly undercutting Sora and Veo. H3 accepts up to... og:url: https://daily.dev/posts/minimax-h3-is-open-weight-2k-video-and-70-cheaper-than-the-competition-ppgcxunyx og:image: https://api.daily.dev/og/posts/pPGcxuNyX.png og:image:alt: MiniMax H3 is open-weight, 2K video, and 70% cheaper than the competition og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # MiniMax H3 is open-weight, 2K video, and 70% cheaper than the competition **[Trends](https://daily.dev/sources/trends)** · 2 min read · 6 upvotes · 1 comments ## Summary MiniMax released H3, an open-weight video generation model priced at $0.081 per 15-second 2K clip — significantly undercutting Sora and Veo. H3 accepts up to 12 multimodal inputs (text, images, audio, video) in a single context window and outputs H.264 video with synchronized stereo audio at up to 2K/24FPS. The vLLM team shipped day-0 support via vLLM-Omni with an OpenAI-compatible endpoint. Weights are available on Hugging Face. The key differentiator beyond price is the open-weight nature: unlike Sora and Veo, H3 can be self-hosted, fine-tuned, and deployed independently. Quality relative to closed competitors remains to be validated by community testing. ## Content MiniMax just dropped H3, an open-weights video generation model that's making the Sora and Veo crowd look expensive. The headline number: $0.081 per 15-second 2K clip. For comparison, Google Veo and OpenAI's Sora aren't in the same zip code on pricing. The model itself is genuinely capable. It takes up to 12 multimodal inputs — text, images, audio, video — in a single unified context and spits out H.264 video with synchronized stereo audio already baked in. No separate audio step. First-and-last-frame control, reference-based motion transfer, consistent lip sync. The full omni-modal package. The ecosystem response was immediate. vLLM shipped day-0 support through vLLM-Omni, serving H3 over an OpenAI-compatible `/v1/videos` endpoint. ComfyUI also had same-day support, and their team did something impressive: through modulation weight pruning, int8 quantization, and custom VRAM offloading, they cut memory requirements from 123.6 GB down to 42.5 GB — a 66% reduction that lets the model run on an RTX 3060. That's a consumer GPU. Simon Willison ran it on an M5 Pro Mac ("a rainbow colored skunk leaps over a mossy log in a supermarket"), though the ~115 GB download and 45-minute generation time are real friction points. Hugging Face's multimodalart called it "powerful" on release. Clement Delangue retweeted the launch. The weights are publicly available on Hugging Face, and platforms like Halo AI and TopviewAI are already offering it — TopviewAI claiming it's 70% cheaper than Seedance. The open-weights angle is what makes this land differently than another API-only model drop. You can run it, modify it, serve it yourself. The community tooling materialized within hours. That's the kind of release that ages well. ## Community take How the wider developer community reacted, aggregated from 1 discussion and 81 comments across hackernews (as of 2026-08-04). **TL;DR:** The community is genuinely impressed by MiniMax H3's quality and the memory optimization work, with many users sharing real benchmark times on their own hardware. There's healthy skepticism about licensing restrictions, quality compared to closed-source alternatives like Seedance 2.5, and the usual philosophical debates about AI's impact on creative industries. **Sentiment:** 55% positive · 35% mixed · 10% skeptical **The case for** - The 66% memory reduction via modulation weight pruning is technically elegant and reportedly lossless, enabling local runs on consumer GPUs like an RTX 3060. - Generation speeds are already usable on mid-range cards (e.g., ~10 min for 10s 480p on a 4070 Ti Super) and expected to improve quickly with sageattention and future distillation. - Being open-weights is seen as a major win, preventing runaway cloud costs and platform safety over-censorship for creative work. - The mouse demo and several other clips were called out as a genuine quality leap over current SOTA open models. - Reference-based motion transfer (first-and-last-frame control) is highlighted as a key missing piece for stitching independent cinematographic scenes. **The pushback** - Licensing is restrictive for users in the US, EU, UK, and South Korea, requiring explicit approval for commercial use. - Quality is disputed — at least one heavy user claims it's still roughly a year or more behind closed-source leaders like Seedance 2.0/2.5 and Kling. - Output aesthetics are described by some as bland, generic, and exhibiting the typical 'AI smoothening' effect. - The model doesn't reliably follow complex prompt instructions (e.g., specific camera transitions were simply ignored). - Mac support appears broken at launch, and very high VRAM configurations (~83 GB) are needed for full-precision weights. **By community** - hackernews (positive): Commenters are enthusiastic about the technical achievement and are actively benchmarking it on their own hardware, with debate focused on generation speed, licensing, and quality relative to closed-source alternatives. **Hottest debate:** Whether MiniMax H3 is genuinely competitive with closed-source leaders like Seedance 2.5, with one commenter claiming it's 'a year and a half behind' while another cites leaderboard data showing H3 ahead of Seedance 2.0. **Open questions** - How long will generation take at full 2K resolution on consumer hardware, and is the current ComfyUI implementation fully optimized? - Will the restrictive regional licensing effectively block most Western commercial users, and can they get approval? - How does MiniMax H3 actually compare to Seedance 2.5 in head-to-head user preference tests? - Will Mac/Apple Silicon support be fixed, and what performance can be expected on devices like Mac Studio Ultra? - Can the modulation weight pruning / LUT approach be applied to other diffusion model architectures beyond this one? **Highlights** > It is a well-known trick, given that the timestep is between 0 to 1, you can slicing them at any resolution (1000, or 10000, give or take), and then keep a look-up table for modulation scale / bias etc for each. It is quite different from quantization and it is indeed lossless. It is also only applicable to diffusion models as only these operates at per-timestep. > — [liuliu on hackernews · 1 comments](https://news.ycombinator.com/item?id=49158433) > > A year and a half behind Seedance 2.0? That is a bold claim that needs evidence. SOTA a year ago was Kling 2.5, and H3 does not look or perform at that level. The Artificial Analysis rankings are whack. They rank Omni first, which is incredulously wrong. Google's models broadly suck, and there they all are - right at the top. Artificial Analysis has notoriously ranked models such as Grok Imagine highly and continues to rank Happy Horse as a good model despite the model being absolute garbage. Could it be because they are subject to broad based statistical attacks? It's easy to encode information about the origin of media in either its metadata or output frames. Or maybe there's simply no overlap between creatives and people who click on ELO scores. I've spent thousands upon thousands of dollars generating video. I will stand by the claim that nothing touches Seedance 2.0 / 2.5 It's good that we're getting better open weights, but these are not a substitute for Kling or Seedance yet. > — [echelon on hackernews](https://news.ycombinator.com/item?id=49161557) > I mean, even the demo prompts on the ComfyUI page aren't adhered to by the model. From the first prompt, one of the four lines: > TRANSITION: a violent WHIP PAN off the rooftop that SMEARS the floating words away with it, motion-streaked — And the video just didn't do any of that transition at all, it just replaced it with a cut. If you look at the rest of the prompts, you'll find similar lines that are just totally ignored. Except maybe the mouse one, I didn't see anything wrong with that off the bat. > — [fwip on hackernews](https://news.ycombinator.com/item?id=49156863) > This is still about a year and a half behind Seedance 2.0/ Seedance 2.5 But it represents a coming price pressure that will face the leading foundation models. Open source will prevent runaway costs. Moreover, it prevents the hair-trigger platform safety checkers from shutting down creative work. Video models are notoriously bad at shutting down a huge number of requests. Creatives will prefer to work on cloud or private GPU clusters. Waiting 10 minutes for a few seconds of 480p is unacceptable. Hobbyists will have fun, but most actual production work is happening in the cloud. Artist's time is worth money, and they like to spin up dozens of concurrent generations at a time to more quickly explore the generation state space and make progress on completing work. > — [echelon on hackernews · 1 comments](https://news.ycombinator.com/item?id=49156651) > Depends on what weights, I landed on: diffusion model: minimax_h3_fl2va_bf16.safetensors text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors video VAE: minimax_h3_video_vae_fp16.safetensors audio VAE: minimax_h3_audio_vae_fp32.safetensors Ends up at ~83GB, but they also shipped bunch of pruned + quantized versions of the diffusion model, might fit with a 5090: https://huggingface.co/Comfy-Org/MiniMax-H3 > — [embedding-shape on hackernews](https://news.ycombinator.com/item?id=49160981) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49155629) · 245 points · 81 comments ## Community discussion Top comments from developers on daily.dev. **@petecapecod** · 0 upvotes > A good oss video model, I mean Seeddance does make incredible videos, be cool to see what Mm3 can do --- Tags: [#open-source](https://daily.dev/tags/open-source), [#genai](https://daily.dev/tags/genai), [#video-generation](https://daily.dev/tags/video-generation), [#vllm](https://daily.dev/tags/vllm) [View this post on daily.dev](https://daily.dev/posts/minimax-h3-is-open-weight-2k-video-and-70-cheaper-than-the-competition-ppgcxunyx) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"MiniMax H3 is open-weight, 2K video, and 70% cheaper than the competition","url":"https://daily.dev/posts/minimax-h3-is-open-weight-2k-video-and-70-cheaper-than-the-competition-ppgcxunyx","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/minimax-h3-is-open-weight-2k-video-and-70-cheaper-than-the-competition-ppgcxunyx"},"datePublished":"2026-08-03T02:49:06.702Z","dateModified":"2026-08-04T19:13:52.041Z","description":"MiniMax released H3, an open-weight video generation model priced at $0.081 per 15-second 2K clip — significantly undercutting Sora and Veo. H3 accepts up to...","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/minimax-h3-is-open-weight-2k-video-and-70-cheaper-than-the-competition-ppgcxunyx","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":6},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"open-source,genai,video-generation,vllm","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"MiniMax H3 is open-weight, 2K video, and 70% cheaper than the competition"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/minimax-h3-is-open-weight-2k-video-and-70-cheaper-than-the-competition-ppgcxunyx","comment":[{"@type":"Comment","text":"A good oss video model, I mean Seeddance does make incredible videos, be cool to see what Mm3 can do","datePublished":"2026-08-03T12:53:43.350Z","url":"https://daily.dev/posts/pPGcxuNyX#c-xk98j8Wa5","author":{"@type":"Person","name":"Peter Cruckshank","url":"https://daily.dev/petecapecod","image":"https://media.daily.dev/image/upload/s--ZJhQyKws--/f_auto/v1721235024/avatars/avatar_A9xh33q0QoxtkGoJRCosp"}}]} ```