<h2 id="support">Support</h2><p>Everything I publish is free and stays free. If it saved you a night of debugging, tips keep the 5090 warm: <a target="_blank" rel="ugc" href="https://ko-fi.com/joeygambino">Ko-fi</a> · <a target="_blank" rel="ugc" href="https://github.com/sponsors/jlucasmcrell">GitHub Sponsors</a> · <a target="_blank" rel="ugc" href="https://liberapay.com/joeygambino">Liberapay</a>.</p><p></p><p><strong>The text encoder half of MiniMax-H3, quantized.</strong> H3 runs on two separate models: the 33B video+audio DiT, and this — a <strong>Qwen3-VL-32B vision-language encoder</strong>. The DiT on its own will not generate anything. If you downloaded the H3 GGUF weights and had nowhere to point the CLIP loader, this is the missing file.</p><h2 id="the-files">The files</h2><ul><li><p><strong>Q4_K_M</strong> — 19.8 GB · about 16.5 GB resident. <strong>Recommended.</strong></p></li><li><p><strong>Q5_K_M</strong> — 23.2 GB · higher precision, more VRAM.</p></li><li><p><strong>mmproj F16</strong> — 1.2 GB · the multimodal projector. <strong>Download this one too.</strong> See below.</p></li></ul><p>Yes, these are real K-quants — unlike the H3 DiT, which cannot have them. K-quants need weight rows divisible by 256, and the DiT's hidden width is 2688, which is not. This encoder is a stock Qwen3-VL and its rows are, so the whole K-ladder is available here.</p><h2 id="the-mmproj-is-not-optional-and-it-pairs-by-filename">The mmproj is not optional — and it pairs by filename</h2><p>The mmproj is the encoder's <strong>vision path</strong>. Everything that conditions on an image goes through it: image-to-video, reference-to-video, and <strong>multi-shot chaining</strong> — chaining works by feeding the previous shot's last frame back through the encoder as an image, so it needs the projector even when you never attached a reference picture yourself.</p><p>ComfyUI-GGUF locates it by <strong>matching the encoder's filename</strong>. Rename either file, or keep them in different folders, and pairing silently fails. Text-only prompts keep working, which is exactly what makes this hard to diagnose — it presents as the model ignoring your reference image rather than as a missing file.</p><p>Keep both names exactly as downloaded, in the same folder. The attached setup pack includes <code>check_encoder_setup.py</code>, which tells you in one line whether the pairing resolves.</p><h2 id="install">Install</h2><pre><code>ComfyUI/models/text_encoders/
  MiniMax-H3-encoder-Q4_K_M.gguf
  MiniMax-H3-encoder-mmproj-F16.gguf</code></pre><p>Load with <code>CLIPLoaderGGUF</code> from <a target="_blank" rel="ugc" href="https://github.com/city96/ComfyUI-GGUF">ComfyUI-GGUF</a>, or with <code>H3ClipLoaderAny</code> from <a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/ComfyUI-H3-Multishot">ComfyUI-H3-Multishot</a>, which takes <code>.safetensors</code> or <code>.gguf</code> in one dropdown. Set the CLIP type to <code>minimax</code>.</p><p><strong>No architecture patch is needed for the encoder.</strong> The H3 <em>DiT</em> needs <code>minimax_h3</code> added to ComfyUI-GGUF's image-architecture list; this file is Qwen3-VL, and <code>qwen3vl</code> is already in its text list. Nothing to run.</p><h2 id="vram-read-this-before-reporting-a-slow-render">VRAM — read this before reporting a slow render</h2><p>The encoder (~16.5 GB) and the H3 DiT (~25 GB) <strong>do not co-fit on a 32 GB card</strong>. If both are held at once, the DiT loads <strong>partially</strong> and streams the remainder from system RAM on every sampling step. Measured on an RTX 5090: <strong>~60 minutes</strong> for a clip that takes <strong>~15 minutes</strong> with the DiT fully resident. The tell in your log is:</p><pre><code>loaded partially; 6423 MB usable, 5847 MB loaded, 19363 MB offloaded</code></pre><p>The fix is to evict the encoder after conditioning and before sampling. The samplers in ComfyUI-H3-Multishot do it automatically and print <code>[H3Multishot] TE evicted; NN.N GB free for the DiT</code>.</p><h2 id="what-else-you-need-to-run-h3">What else you need to run H3</h2><ul><li><p>MiniMax-H3 GGUF — the DiT (<code>fl2va</code> or <code>ref2va</code>), on <a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/MiniMax-H3-GGUF">Hugging Face</a> and in my Civitai listings.</p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3">Comfy-Org/MiniMax-H3</a> — the video and audio VAEs.</p></li><li><p><a target="_blank" rel="ugc" href="https://github.com/jlucasmcrell/ComfyUI-H3-Multishot">ComfyUI-H3-Multishot</a> — multi-shot, I2V, and the encoder eviction above.</p></li></ul><h2 id="everything-else-i've-published">Everything else I've published</h2><ul><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/MiniMax-H3-encoder-GGUF">This encoder on Hugging Face</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/MiniMax-H3-GGUF">H3 DiT weights on Hugging Face</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino/MiniMax-H3-Multishot-Workflow">The multishot workflow</a></p></li><li><p><a target="_blank" rel="ugc" href="https://huggingface.co/joeygambino">All models on Hugging Face</a></p></li></ul><h2 id="support">Support</h2><p>Everything I publish is free and stays free. If it saved you a night of debugging, tips keep the 5090 warm: <a target="_blank" rel="ugc" href="https://ko-fi.com/joeygambino">Ko-fi</a> · <a target="_blank" rel="ugc" href="https://github.com/sponsors/jlucasmcrell">GitHub Sponsors</a> · <a target="_blank" rel="ugc" href="https://liberapay.com/joeygambino">Liberapay</a>.</p>

Files with SHA256: 448b7ab3ca0b1f92a5ebca9cc6a4b56e7dbb45b0e313c9d34d8a1154eb49d7ea