--- title: Show HN: We compressed a 31B video+audio diffusion model... description: HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of MiniMax-H3 that generates video with synchronized stereo audio. It runs in 4, 6, or 8... canonical: https://daily.dev/posts/show-hn-we-compressed-a-31b-video-audio-diffusion-model-to-14b-6-steps-int8-1p77bmj76 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8 | daily.dev og:description: HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of MiniMax-H3 that generates video with synchronized stereo audio. It runs in 4, 6, or 8... og:url: https://daily.dev/posts/show-hn-we-compressed-a-31b-video-audio-diffusion-model-to-14b-6-steps-int8-1p77bmj76 og:image: https://api.daily.dev/og/posts/1p77BMj76.png og:image:alt: Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8 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. # Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8 **[Hacker News](https://daily.dev/sources/hn)** · 17 min read · 1 upvotes · 0 comments ## Summary HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of MiniMax-H3 that generates video with synchronized stereo audio. It runs in 4, 6, or 8 sampling steps without classifier-free guidance, producing 5.2-15 second clips at 24 fps and 1344x768 resolution by default with 32 kHz stereo audio. The model has 13.7B to 13.9B parameters, less than half the ~31B MiniMax-H3 transformer, and runs as GGUF (as small as 7.2 GB in Q3_K_M) or INT8 convrot quantization. A 5.2 second clip takes 8.7s on 4x B200 GPUs, 36s on 1x H200, or 92s on a single RTX 5090. NVIDIA's SOL-H3 engine runs it in real time on 4x B-series GPUs. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How fast can HyperFlow H3 Flash generate a 5 second video clip on an RTX 5090? A 5.2 second clip at 1344x768 with 6 sampling steps takes about 92 seconds on a single RTX 5090 using the INT8 pipeline, or 99 to 130 seconds with GGUF depending on system RAM. On data-center hardware it drops to 36 seconds on one H200 and 8.7 seconds on 4x B200 GPUs. _daily.dev surfaces benchmark data like this for developers sizing GPU hardware for video generation workloads._ ### What is the minimum VRAM needed to run HyperFlow H3 Flash locally? The Q4_K_M GGUF quantization needs 9.1 GB and runs on 24 GB or 32 GB consumer GPUs with as little as 16 GB of system RAM. The INT8 convrot variant needs roughly 16 GB of VRAM but requires at least 69 GiB of system RAM when offloading with an fp8 text encoder, or 91 GiB with bf16. _Developers picking a runtime for local generative video models can track setup tradeoffs like these on daily.dev._ ### What is the difference between GGUF and INT8 convrot quantization for this model? GGUF quantization (Q3_K_M through Q8_0, 7.2 to 15.2 GB) targets consumer RTX GPUs with limited system RAM, running through ComfyUI, Unsloth Studio, or stable-diffusion.cpp. INT8 convrot quantization uses W8A8 precision with low-rank factorization, runs 10 to 20 percent faster on data-center GPUs or 24GB+ RTX cards, but needs far more system RAM, around 69 to 91 GiB with offloading. _Comparing quantization formats before deployment is easier when developers track benchmarks like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 24 comments across hackernews (as of 2026-11-08). **TL;DR:** Developers are intrigued by the compression technique and practical run times across hardware, but many want head-to-head quality comparisons across quantization levels before judging it a success. **Sentiment:** 35% positive · 55% mixed · 10% skeptical **The case for** - The author's detailed technical breakdown (pruning, low-rank factorization, INT8 with Hadamard rotation) was well received. - Some found it genuinely useful to run locally via stable-diffusion.cpp style CLI workflows. - Practical performance numbers across different GPUs were appreciated. **The pushback** - Several commenters want direct quality comparisons between quantization levels and the original model before trusting the compression claims. - One commenter compared it to compressing a JPEG without checking what the result looks like, questioning the lack of quality evaluation. - ComfyUI is seen as cumbersome, with requests for a simpler CLI-based workflow. **By community** - hackernews (mixed): Interest in the technique and tooling is tempered by repeated requests for quality-vs-quantization comparisons that aren't yet provided. **Open questions** - How does quality degrade across the different quantization levels (GGUF Q3_K_M vs Q4_K_M vs INT8) compared to the original 31B model? - How does this compare in perf/quality to other compressed alternatives like FastVideo's FastH3? - What are performance numbers on 16GB GPUs? **Highlights** > Hi HN, I’m Jian, one of the engineers on this. MiniMax-H3 is an open model that generates video and audio together, but its transformer is ~31B parameters and sampling takes a long trajectory, so it’s slow and heavy to run locally. HyperFlow H3 Flash is our compressed version. What we changed: Pruning + low-rank factorization: dropped 3 of 50 blocks and factored 102 linear layers, down to 13.7B params. Two-time (t, r) conditioning: each step knows both the current noise level and where it should land, so it samples in 4–8 steps without CFG. W8A8 INT8 with a Hadamard rotation on 180 linear layers. End-to-end numbers for a 5.2 s clip at 1344×768, 6 steps: 8.7 s on 4× B200, 36 s on 1× H200, 92 s on an RTX 5090, 266 s on an RTX 3090. The Q4_K_M GGUF is 9.1 GB and runs in ComfyUI on 24 GB cards. 16 GB GPUs aren’t measured yet. To be clear about “real time”: our package does 8.7 s for a 5.2 s clip on 4× B200. Real-time playback speed uses NVIDIA’s SOL-H3 runtime, which supports the model from day 0. License: code is Apache-2.0 (ComfyUI nodes GPL-3.0). The weights are a derivative of MiniMax-H3 under its Community License, which does not cover the US, EU, UK or South Korea without MiniMax’s authorization. Please check before using. Demos and an overview: https://www.videorebirth.com/lp/hyperflow-flash Happy to answer questions about the compression or the numbers. > — [jiant on hackernews](https://news.ycombinator.com/item?id=50405264) > Yeah this is only meaningful if you can compare quality at least. Otherwise, it's just like saying: we compressed a JPEG file to 50% of its size using some compression level and we didn't check what it looks like. > — [amelius on hackernews](https://news.ycombinator.com/item?id=50407410) > Seconded, I'd like to see a head to head comparison to see how the quant loks vs the original > — [Reubend on hackernews](https://news.ycombinator.com/item?id=50406656) > Am curious how this compares perf/quality wise with FastVideo's FastH3 trim/omni which I have been using pretty well with a basic laptop NVidia card (e.g. 5s 480p video w/ audio in 90s). > — [kodablah on hackernews](https://news.ycombinator.com/item?id=50407208) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50405263) · 56 points · 24 comments ## Similar posts on daily.dev - [MiniMax H3 on vLLM-Omni: From System-Wide Optimization to Real-Time Serving with FastVideo’s FastH3](https://daily.dev/posts/minimax-h3-on-vllm-omni-from-system-wide-optimization-to-real-time-serving-with-fastvideo-s-fasth3-mhkh8zrjp) · vLLM · 0 upvotes · 0 comments - [Run Step 3.7 Flash on NVIDIA GPUs with Enterprise-Ready Multimodal AI](https://daily.dev/posts/run-step-3-7-flash-on-nvidia-gpus-with-enterprise-ready-multimodal-ai-9aofwl0f4) · NVIDIA Developer · 1 upvotes · 0 comments --- Tags: [#python](https://daily.dev/tags/python), [#local-ai](https://daily.dev/tags/local-ai) [View this post on daily.dev](https://daily.dev/posts/show-hn-we-compressed-a-31b-video-audio-diffusion-model-to-14b-6-steps-int8-1p77bmj76) ```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":"Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8","url":"https://daily.dev/posts/show-hn-we-compressed-a-31b-video-audio-diffusion-model-to-14b-6-steps-int8-1p77bmj76","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/show-hn-we-compressed-a-31b-video-audio-diffusion-model-to-14b-6-steps-int8-1p77bmj76"},"datePublished":"2026-11-08T00:03:46.435Z","dateModified":"2026-11-08T00:04:24.791Z","description":"HyperFlow H3 Flash is Video Rebirth's compressed, few-step version of MiniMax-H3 that generates video with synchronized stereo audio. It runs in 4, 6, or 8...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ecead293dd62f23bfc1df3a220fa63fe?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ecead293dd62f23bfc1df3a220fa63fe?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","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":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/show-hn-we-compressed-a-31b-video-audio-diffusion-model-to-14b-6-steps-int8-1p77bmj76","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"python,local-ai","timeRequired":"PT17M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Show HN: We compressed a 31B video+audio diffusion model to 14B, 6 steps, INT8"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/show-hn-we-compressed-a-31b-video-audio-diffusion-model-to-14b-6-steps-int8-1p77bmj76#faq","mainEntity":[{"@type":"Question","name":"How fast can HyperFlow H3 Flash generate a 5 second video clip on an RTX 5090?","acceptedAnswer":{"@type":"Answer","text":"A 5.2 second clip at 1344x768 with 6 sampling steps takes about 92 seconds on a single RTX 5090 using the INT8 pipeline, or 99 to 130 seconds with GGUF depending on system RAM. On data-center hardware it drops to 36 seconds on one H200 and 8.7 seconds on 4x B200 GPUs. daily.dev surfaces benchmark data like this for developers sizing GPU hardware for video generation workloads."}},{"@type":"Question","name":"What is the minimum VRAM needed to run HyperFlow H3 Flash locally?","acceptedAnswer":{"@type":"Answer","text":"The Q4_K_M GGUF quantization needs 9.1 GB and runs on 24 GB or 32 GB consumer GPUs with as little as 16 GB of system RAM. The INT8 convrot variant needs roughly 16 GB of VRAM but requires at least 69 GiB of system RAM when offloading with an fp8 text encoder, or 91 GiB with bf16. Developers picking a runtime for local generative video models can track setup tradeoffs like these on daily.dev."}},{"@type":"Question","name":"What is the difference between GGUF and INT8 convrot quantization for this model?","acceptedAnswer":{"@type":"Answer","text":"GGUF quantization (Q3_K_M through Q8_0, 7.2 to 15.2 GB) targets consumer RTX GPUs with limited system RAM, running through ComfyUI, Unsloth Studio, or stable-diffusion.cpp. INT8 convrot quantization uses W8A8 precision with low-rank factorization, runs 10 to 20 percent faster on data-center GPUs or 24GB+ RTX cards, but needs far more system RAM, around 69 to 91 GiB with offloading. Comparing quantization formats before deployment is easier when developers track benchmarks like this on daily.dev."}}]} ```