<h1 id="minimax-h3-pruned-fp8-scaled">MiniMax H3 pruned fp8 scaled</h1>Pruned FP8 version of the video model<h3 id="fl2va:">fl2va:</h3>(I2V, FL2VA, T2V)Direct link: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_fp8_scaled.safetensors?download=true">https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_fp8_scaled.safetensors?download=true</a><h3 id="ref2va:">ref2va:</h3>Direct link: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_fp8_scaled.safetensors?download=true">https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_fp8_scaled.safetensors?download=true</a>-Other required files:nvfp4 Text Encoder: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors?download=true">https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors?download=true</a>Video VAE: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors?download=true">https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors?download=true</a>Audio VAE: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors?download=true">https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors?download=true</a>

MiniMax H3 pruned fp8 scaled

MiniMax H3 fl2va pruned fp8 scaled

minimax_h3_video_vae_fp16.safetensors

minimaxH3PrunedFp8_fl2vaPrunedFp8Scaled_txt.safetensors

minimaxH3PrunedFp8_fl2vaPrunedFp8Scaled.safetensors

minimax_h3_audio_vae_fp32.safetensors

fl2va pruned fp8 scaled

ref2va pruned fp8 scaled

Minimax H3 INT8 / INT6 / INT4 ConvRotSampler settings I currently use are: er_sde/res_multistep simpleQuantized weights of the Minimax H3 video generation model (FL2VA and REF2VA variants), packaged for ComfyUI. Four precision tiers are available so you can pick the right point on the speed / quality / VRAM/RAM curve for your hardware.<h3 id="about-these-variants">About these variants</h3>Four precision levels are available for both FL2VA and REF2VA (all sizes are pruned checkpoints), trading VRAM and speed against generation quality:<ul><li>BF16 — full-precision reference (~40 GB). Highest quality, largest footprint. Use if you have the VRAM and want the ceiling to compare against.  </li><li>INT8 ConvRot (W8A8) — 8-bit weights and 8-bit activations (~21 GB). Near-lossless quality, runs on INT8 tensor cores (RTX 30-series and up). The default recommendation.  </li><li>INT6 (W6A8) — 6-bit weights with 8-bit activations (~16 GB). Middle-ground between INT8 and INT4: noticeably smaller than INT8 with a small quality trade-off. Good pick when INT8 is close to fitting but not quite. Uses the W6A8 kernel path from <a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/comfy-kitchen/pull/191">comfy-kitchen PR #191</a>.  </li><li>INT4 (W4A8 mixed) — 4-bit weights with 8-bit activations (~12.5 GB), quantized by Kijai. Smallest footprint of the group. Uses a ConvRot codebook layout (<a target="_blank" rel="ugc" href="https://github.com/Comfy-Org/comfy-kitchen/pull/90">comfy-kitchen #90</a>) that decodes to INT8 at compute time, so it runs on the same INT8 hardware as W8A8 with the same throughput — the win is memory, not speed. Larger quality trade-off than INT6.</li></ul>All quantized variants keep the sensitive layers (I/O, modulation, embeddings, output head) in BF16. If you're not sure: start with INT8. Move to INT6 if you need a bit more room. Move to INT4 if you're actively hitting VRAM limits — the quality gap to INT8 (and INT8 to INT6) is small enough that it's a good tradeoff on tight budgets.  <h3 id="model-variants">Model variants</h3>FL2VA — first-last frame to video / audio  REF2VA — reference-conditioned: up to 9 reference images, 3 reference videos (each may carry its own paired soundtrack), and 3 standalone reference audio clips   Both models support t2v (text to video), i2v (image to video), v2v (video to video), a2v (audio to video), and multiple mixed references (image / video / audio). Both were further fine-tuned for higher-quality outputs in their intended use cases — FL2VA for first-and-last-frame interpolation, REF2VA for reference-driven generation.  <h3 id="requirements">Requirements</h3><ul><li>ComfyUI at a version that supports Minimax H3 natively.  </li><li>For the INT4 (W4A8) model: ComfyUI v0.31.0 or newer (PR #15308).  </li><li>An NVIDIA GPU with useful INT8 throughput (RTX 30-series and up recommended).  </li></ul>

Minimax H3 INT8/INT6/INT4 ConvRot

minimaxH3INT8INT6INT4_fl2vaINT8Pruned.safetensors

minimaxH3INT8INT4_fl2vaINT8Pruned_txt.safetensors

minimaxH3INT8INT6INT4_fl2vaINT8Pruned.json

FL2VA INT8 Pruned

FL/REF2VA Pruned INT6W6A8

FL/REF2VA Pruned AsymW4A8

REF2VA INT4BQ Pruned

REF2VA INT4Q DEPRECATED

FL2VA INT4Q DEPRECATED

FL2VA INT4BQ Pruned

REF2VA INT8 Pruned

<a target="_blank" rel="ugc" href="https://hailuoai.video/">Hailuo H3</a> is the newest video generation model from <a target="_blank" rel="ugc" href="https://www.minimax.io/">MiniMax</a>, the Shanghai AI lab behind the Hailuo video app and the MiniMax text and audio model families. H3 generates native 2K video with synchronized audio in a single pass, at clip lengths up to 15 seconds. Developed by MiniMax. All credit for the model goes to the MiniMax team. This page carries four versions of H3:• Comfy - the open weights. Generate with them on-site (the only version that takes LoRAs), or download them and run them yourself. The ComfyUI-format weights are also available from MiniMax and Comfy-Org on Hugging Face. • Api - H3 through the hosted <a target="_blank" rel="ugc" href="https://platform.minimax.io/docs/guides/video-generation">MiniMax API</a>, at native 2K. • Max - MiniMax's H3 Max release, with a faster Max Turbo option, at 480p, 768p, or 1080p. • HeyGen - <a target="_blank" rel="ugc" href="https://www.heygen.com/">HeyGen</a> Video 1, HeyGen's model built on H3, at 480p or 768p. Credit for this version goes to the HeyGen team.<h3 id="background">Background</h3>MiniMax has shipped video models under the Hailuo name since 2024, moving from the I2V-01 series through Hailuo 02 and Hailuo 2.3. Those earlier models topped out around 1080p with fixed 6 or 10 second clips and no native sound. H3 is the step past that: it raises the ceiling to 2K, opens the clip length to any integer from 4 to 15 seconds, and folds audio generation into the same pass as the picture rather than bolting it on afterward.<h3 id="capabilities">Capabilities</h3>• Native 2K output (2560x1440 at 16:9) • Clips from 4 to 15 seconds, in one second increments • Synchronized audio generated with the video - dialogue, effects, and ambience timed to what is on screen • Text-to-video, image-to-video, and first frame plus last frame control • Omni-reference conditioning: up to 9 reference images, 3 reference video clips, and 3 reference audio clips • Multi-shot sequences and instruction-based editing<h3 id="on-civitai">On Civitai</h3>H3 is wired into the video generator across four workflows: Create Video (text only), Image to Video (single first frame), First/Last Frame (both endpoints), and Reference to Video (up to 9 reference images). Pick the version from the model selector:• Api - resolution is fixed at 2K, the only output this version produces. • Max - choose 480p, 768p, or 1080p, and switch on Turbo for faster Max Turbo generations. Set a seed to repeat a result. • HeyGen - choose 480p or 768p, with a seed. It takes a first frame but no last frame, so it is not offered under First/Last Frame. • Comfy - runs the open weights and takes LoRAs, with control over steps and a turbo LoRA.Aspect ratio is selectable at 16:9, 4:3, 1:1, 3:4, and 9:16 for text prompts and reference images; when you supply a frame image, the framing follows that image instead.Reference video and reference audio inputs are supported by the models but are not exposed in the Civitai form yet.<h3 id="notes">Notes</h3>H3 is a new model and MiniMax has not published a full spec sheet or a stable pricing page for it. Behavior and cost may shift as they finish the rollout. The Api, Max, and HeyGen versions are commercially hosted services, and content is subject to each provider's policies, so expect refusals on material their APIs decline.<h3 id="links">Links</h3>• <a target="_blank" rel="ugc" href="https://platform.minimax.io/docs/guides/video-generation">MiniMax video generation docs</a> • <a target="_blank" rel="ugc" href="https://hailuoai.video/">Hailuo AI</a> • <a target="_blank" rel="ugc" href="https://www.minimax.io/">MiniMax</a> • <a target="_blank" rel="ugc" href="https://www.heygen.com/">HeyGen</a> • <a target="_blank" rel="ugc" href="https://hailuoai.video/doc/terms-of-service.html">MiniMax Hailuo AI Terms of Service</a>

Files with SHA256: 35a88d51044231fe332301d7a62aa81e3f2cba62febeb446e2c1e3e0ef76f2c6