<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Claire Mitchell on Medium]]></title>
        <description><![CDATA[Stories by Claire Mitchell on Medium]]></description>
        <link>https://medium.com/@claire_mitchell?source=rss-15039f8d3a79------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ZHrn92Dx4E-39WA39nulvQ.png</url>
            <title>Stories by Claire Mitchell on Medium</title>
            <link>https://medium.com/@claire_mitchell?source=rss-15039f8d3a79------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:45:11 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@claire_mitchell/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[MiniMax H3 Locally Is Possible. It’s Also a 31.7GB VRAM Problem]]></title>
            <link>https://medium.com/@claire_mitchell/minimax-h3-locally-is-possible-its-also-a-31-7gb-vram-problem-108b9f12d5c0?source=rss-15039f8d3a79------2</link>
            <guid isPermaLink="false">https://medium.com/p/108b9f12d5c0</guid>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[video-generation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[minimax]]></category>
            <category><![CDATA[comfy-ui]]></category>
            <dc:creator><![CDATA[Claire Mitchell]]></dc:creator>
            <pubDate>Wed, 12 Aug 2026 13:58:11 GMT</pubDate>
            <atom:updated>2026-08-12T13:58:11.778Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UZg0FgdnIryFLtXYhxPurw@2x.png" /></figure><p>Every time a big video model goes open source, the same question floods the timeline within hours: “Can I run this?” MiniMax H3 — also known as Hailuo 3.0 — landed as open weights at the end of July 2026, and ComfyUI support followed within days. So the conversation quickly shifted from “does it run at all” to “what does running it well actually cost.” That’s the question I went down a rabbit hole on this week, and the answers are not what the spec sheets suggest.</p><p>Full disclosure up front: I’m working from a detailed local benchmark that someone ran on an RTX 5090–36 videos across three checkpoints, three prompts, and four resolutions, roughly 15,000 seconds of compute — rather than from my own four-GPU rig (I wish). I’ll link the full writeup at the end, because it’s the kind of evidence-first testing this space doesn’t get enough of. But first, the part that tripped everyone up.</p><p><strong>The trap: file size isn’t VRAM.</strong></p><p>H3 ships two open-weight models: fl2va, which handles text-to-video and image-to-video, and ref2va, which is the reference-to-video variant. The pruned INT8 checkpoint is about 21GB on disk. If you’re like me, your brain immediately does the math — “21GB, I have a 24GB card, we’re good.” We are not good.</p><p>The benchmark measured peak VRAM at roughly 31.7GB for every checkpoint it tested: full INT8 (~32GB on disk), pruned INT8 (~21GB), and pruned FP8 (~21GB). All three, same ceiling. That’s basically maxing out the RTX 5090’s 32GB, regardless of how small the download was.</p><p>What pruning actually buys you is system RAM, not VRAM. The full INT8 run peaked at 93.1GB of system memory, pruned INT8 at 77.6GB, and pruned FP8 at 71.2GB. So the smaller checkpoint is still the right download — it just saves you disk space and RAM pressure. If you were planning your hardware around the file size, now is a good time to redo that math.</p><p><strong>Then there’s the speed surprise.</strong></p><p>They timed everything the honest way: from the moment ComfyUI accepts the job to the finished MP4 sitting in the output folder — prompt processing, diffusion, the audio and video decoding, and the final muxing all included. Every clip is 192 frames at 24fps, which is exactly 8 seconds, rendered at 25 steps with denoise at 1.0. Here’s the full matrix:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/713/1*XhpvhMexcfpCjmMM-A7dgA@2x.png" /></figure><p>Two things jump out. First, pruned INT8 is the obvious default: it’s roughly 23% faster than full INT8 at the lowest resolution, and by 1MP the gap collapses to under 1%. Once resolution becomes the bottleneck, the trimmed weights stop buying you anything — the diffusion and decode work dominate at that point.</p><p>Second, and this is the one nobody called coming: pruned FP8 was the slowest option on this card, about 8–18% slower than pruned INT8 depending on the tier. On a Blackwell GPU, with all the marketing around FP8, that’s genuinely counterintuitive. It’s the same lesson that keeps showing up in the quantized Wan world — smaller file and newer format don’t automatically mean faster, and precision behavior is specific to your GPU generation.</p><p><strong>Which resolution should you actually render?</strong></p><p>Using pruned INT8, the cost ladder is roughly 2 minutes at 0.3MP, 4m17s at 0.5MP, 7m26s at 0.7MP, and 13m49s at 1.0MP. The workflow that fell out of the testing makes sense: draft at 0.3, batch-screen at 0.5, render finals at 0.7, and treat 1.0 as a special occasion. 1.0MP costs about 86% more time than 0.7MP, and what you get is texture — not sense. More pixels give you a sharper version of wrong, not a righter version.</p><p>The test prompts made that painfully clear. The barista sequence — grinding, extracting, pouring latte art — is the classic anatomy nightmare: hands, portafilter, and machine parts fuse together at low resolution, and even at 1MP the object relationships stay shaky. The reef scene, with waves, sunset, and a red sailboat, was the stable one, and 0.7MP looked like its sweet spot for texture and motion. The rainy Tokyo night scene was the most revealing: neon, wet reflections, a cyclist passing, an elevated train appearing on cue. At low resolutions that final train event gets delayed or weakened, and bumping the resolution doesn’t fix the timing — it just adds street detail. It also still renders signage as the usual generative gibberish.</p><p><strong>What “running it well” actually means for your hardware.</strong></p><p>The benchmark rig was an RTX 5090 (32GB), an i9–14900KF, and 128GB of RAM on ComfyUI 0.30.0. And this is not a light workload: near-100% GPU utilization for minutes at a time, roughly 580W of peak draw, and a 91°C peak in the logs. Your power supply, case airflow, and cooling matter just as much as the VRAM here.</p><p>My honest read on the buying advice: 32GB is the tier that just works, and the ~31.7GB peak is exactly why. A used 24GB card like a 3090 can get there with smaller quants and offloading, but it’s a different experience — before handing over money, check memory temperatures, power delivery, and how the card was treated. At 16GB or less, I wouldn’t promise a stable experience at all; a cloud or API workflow is the pragmatic move until the quantization ecosystem matures.</p><p>One fair counterpoint: community GGUF builds have gotten H3 down to an 8.49GB Q2_K quant that technically fits on 8–12GB cards. I don’t doubt it loads. I do doubt it’s a good time — even that writeup’s fine print says “VRAM capacity ≠ comfortable performance,” and MiniMax’s own recommended baseline is four GPUs. “It loaded” is the lowest bar in local AI, and this model clears it the way a lot of models do: technically.</p><p><strong>The fine print, since we’re being honest.</strong></p><p>This is one GPU, one workflow, one seed per case, and three prompts — enough to rank checkpoints and tiers, not enough to characterize output quality in general. No GGUF builds, no first/last-frame or reference-to-video modes, no multi-GPU splits, no offload testing on 16GB or 24GB cards. INT8/FP8 behavior can also shift on Ada or Ampere hardware, so treat these as directional numbers for a high-end single-GPU setup. And if you’re tempted to build something commercial on top: the open weights come with usage restrictions, not a permissive MIT/Apache-style license, so check that before you commit.</p><p>The full benchmark — methodology, raw numbers, the test prompts, the resolution tiers — is here: <a href="https://wan2-7.io/blog/minimax-h3-local-requirements/">MiniMax H3 Local Requirements</a>. It’s worth reading in full if you’re seriously planning a local setup, and it’s a good example of what testing with numbers instead of vibes looks like.</p><p>My takeaway after all of this: H3 locally is genuinely impressive — synced audio, coherent eight-second motion, strong atmosphere. The smart way to run it is also the boring way: grab the ~21GB pruned INT8, render finals at 0.7MP, and stop treating 1MP as a fix-it button. The two mistakes that will cost you hours are assuming file size predicts VRAM, and assuming FP8 is automatically faster. Both are wrong, and now you have the numbers to prove it.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=108b9f12d5c0" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>