# MiniMax H3 Integrations A community-maintained index of checkpoints, tools, and workflows for MiniMax H3, ordered by developer interest. ## Official resources * [MiniMax-H3 official model card](https://huggingface.co/MiniMaxAI/MiniMax-H3) · [official repository](https://github.com/MiniMax-AI/MiniMax-H3) * [Video Prompt Writing Guide — Base (FL2VA)](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md) * [Video Prompt Writing Guide — Reference (Ref2VA)](https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md) * ComfyUI [day-0 blog post](https://blog.comfy.org/p/minimax-h3-day-0-support-in-comfyui) · [tutorial](https://docs.comfy.org/tutorials/video/minimax/minimax-h3) ## Start here This navigation guide is not a complete compatibility list. | Goal | Start with | | :--- | :--- | | Run on your GPU | [Run locally](#models) — pick a stack from the [VRAM table](#recipes-vram) | | Work with audio | [`comfyui-minimax-h3-audio-T8`](https://github.com/T8mars/comfyui-minimax-h3-audio-T8) and [audio VAE](#components-vae) | | Build in ComfyUI | [Official tutorial](https://docs.comfy.org/tutorials/video/minimax/minimax-h3) · [Workflows & nodes](#nodes) | | Write better prompts | [Prompting](#recipes-prompt) | | Make it faster | [Speed](#speed) — [Turbo](#turbo) · [PDD](#pdd) · [FastH3](#fasth3) · [VDN](#vdn) | | Control pose / depth / edges | [Fun Control](#funcontrol) | | Use the hosted model | [Hosted H3 Max](#hosted) | | Fine-tune it | [Training & LoRAs](#training) | | Serve it as an API | [Serving H3](#partners) | | Run on Apple Silicon | [`antirez/h3.c`](https://github.com/antirez/h3.c) | ## Run it locally MiniMax-H3 generates video with **native stereo audio** from text, images, video, and audio inputs. It supports clips up to **2K** and **15 seconds**. Two base variants exist: * **H3-Base-FL2VA** (first-and-last-frame mode) — accepts zero, one, or two input images. Zero images for text-to-video; one image for first- *or* last-frame-to-video; two images for first-and-last-frame-to-video. * **H3-Base-Ref2VA** (omni-reference mode) — accepts up to **9 images, 3 video clips (2–15 s each), and 3 audio clips**, for a maximum of **12 files** total. Checkpoints are the same size. FL2VA, trained only with keyframes, typically yields better raw output. Ref2VA accepts more reference material but has lower base quality; the [Ref Patch](#refpatch) can partially bridge this gap. ### By VRAM and hardware Find your GPU in the table, then use the notes to inform your configuration. | Situation | Stack | Why this combination | | :--- | :--- | :--- | | **24 GB, first run** | `pruned_int8_convrot` DiT (19.53 GiB) + TE `nvfp4_awq` (14.61 GiB) + [`ComfyUI-MiniMaxH3-Easy`](https://github.com/nkxx188/ComfyUI-MiniMaxH3-Easy) | Easy routes T2V, I2V, first/last-frame, and reference input through a single `Media` port. Sampling, LoRAs, and decoding remain outside the node for later modification. | | **24 GB, want speed** | The above + [TE-Speed-MiniMaxH3](https://github.com/tl2012tl/TE-Speed-MiniMaxH3) + Turbo `v4_step600_ema` at **6–8 steps** | Block-cache acceleration; v3.2 adapts to current ComfyUI's block prefetch and no longer patches core files. 6–8 steps reduce Turbo motion smear. | | **22 GB, Turing (sm_75)** | [`minimax-h3-turing`](https://github.com/IvenKooLab/minimax-h3-turing) — W4A8 mixed + Turbo 4-step field handbook | 5.7 min/clip baseline; T8 BlockCache drafts −43%; PDD 8-step + T8 combo 210 s/clip. sm_75 has no FP8/BF16 tensor cores — quantization, launch flags and kernel support all differ from Ampere+ | | **12–16 GB** | Pruned `Q4_K_M` GGUF (10.64 GiB) or pruned `nvfp4` (11.67 GiB) + TE `Q2_K` (7.91 GiB) + fp8mix VAE pair | GGUF offers the most size options, beneficial for tight memory. `IQ1_S` is smaller at 3.78 GiB, but quality noticeably drops. | | **8 GB** | [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio) NF4 path | The project states 8 GB as its minimum for this path. Offloading performs most work here; expect slow performance, not just small memory footprint. | | **RTX 50-series / Blackwell** | [NVIDIA Sol-Attn](https://github.com/kijai/ComfyUI-SolAttn_triton) | **1.14–1.44×** faster than SageAttention with **−37 %** MLP peak VRAM, measured on a 5090. SM89–SM121, Triton 3.6.0. Also unlocks Blackwell-only hybrid-NVFP4 checkpoints. | | **H200 / B200, faster than real time** | [`OpenVDN/vdn-minimax-h3`](https://github.com/OpenVDN/vdn-minimax-h3), hybrid-attention H3 + 8-step adapter | Datacenter path, not a VRAM-saving one: the bf16 H3 base plus a linear-attention branch and two LoRAs. A 768p, 14.4 s clip denoises in **90.5 s on one H200**, **51 s on one B200**, and **11.23 s on 8×B200**. See [VDN](#vdn). | | **Multi-shot / long video** | [`ComfyUI-H3-Motion-Context`](https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context) | H3 generates in blocks up to 15 s. Motion-Context feeds the previous block's final frame **and** audio forward, preserving motion direction and speed. [`Smite79/MiniMax-H3-LongVideos`](https://github.com/Smite79/MiniMax-H3-LongVideos) is a second approach to the same problem, outside ComfyUI. | | **Pose / depth / edge control** | [`alibaba-pai/MiniMax-H3-Fun-Controlnet-Union`](https://huggingface.co/alibaba-pai/MiniMax-H3-Fun-Controlnet-Union) + [`ComfyUI-H3-FunControl`](https://github.com/wyzborrero/ComfyUI-H3-FunControl) | One ~6.8 GiB control branch covers Canny, Depth, HED, MLSD, Pose and video inpainting. See [Fun Control](#funcontrol). | | **Fewest steps** | [PDD 8-step Acc LoRA](#pdd) or [FastH3 4-step preview](#fasth3) | Two newer alternatives to Turbo. PDD is the safer default at 8 steps; FastH3 goes to 4 with more texture loss. | | **Storyboard / timeline** | [`ComfyUI_MiniMaxH3_Director`](https://github.com/huangserva/ComfyUI_MiniMaxH3_Director) | Five importable templates: t2v, fl2v, r2v, v2v, and rv2v. | | **Inpaint / local edit** | [`scraed/LanPaint`](https://github.com/scraed/LanPaint) | v2.1.0 fixed H3 support. Training-free video **and** audio inpainting. | | **Apple Silicon** | [`antirez/h3.c`](https://github.com/antirez/h3.c) (MIT, Metal-native) | h3.c supports T2V/A, first-last-frame, and ordered Ref2VA references end-to-end, with M3 Max / M5 Max performance optimization ongoing. | | **One-command local** | [`open-video-ai/open-video`](https://github.com/open-video-ai/open-video) | "Ollama for video models" — `install` · `pull` · `run`. | ### Checkpoints | Source | What it is | Files | Total | | :--- | :--- | :---: | ---: | | [MiniMaxAI/MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) | Original diffusers weights — `transformer/` (FL2VA) and `transformer_ref/` (Ref2VA) at 14 shards / 61.73 GiB each, plus text encoder, video VAE, audio VAE, and self-contained `FL2VA/` and `Ref2VA/` pipeline folders | 280 | 464.2 GiB | | [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) | ComfyUI-repackaged single-file weights — 10 diffusion models, 3 text encoders, video + audio VAE | 17 | 433.2 GiB | | Variant | Name | Precision | Size | Download | | :--- | :--- | :---: | :---: | :---: | | FL2VA | `minimax_h3_fl2va` | ![bf16][badge-bf16] | 61.73 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_bf16.safetensors) | | FL2VA | `minimax_h3_fl2va` | ![int8][badge-int8] | 31.70 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_int8_convrot.safetensors) | | FL2VA | `minimax_h3_fl2va_pruned` | ![bf16][badge-bf16] | 37.46 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_bf16.safetensors) | | FL2VA | `minimax_h3_fl2va_pruned` | ![fp8][badge-fp8] | 19.52 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_fp8_scaled.safetensors) | | FL2VA | `minimax_h3_fl2va_pruned` | ![int8][badge-int8] | 19.53 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors) | | Ref2VA | `minimax_h3_ref2va` | ![bf16][badge-bf16] | 61.73 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_bf16.safetensors) | | Ref2VA | `minimax_h3_ref2va` | ![int8][badge-int8] | 31.70 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_int8_convrot.safetensors) | | Ref2VA | `minimax_h3_ref2va_pruned` | ![bf16][badge-bf16] | 37.46 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_bf16.safetensors) | | Ref2VA | `minimax_h3_ref2va_pruned` | ![fp8][badge-fp8] | 19.52 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_fp8_scaled.safetensors) | | Ref2VA | `minimax_h3_ref2va_pruned` | ![int8][badge-int8] | 19.53 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors) | **"Pruned"** means AdaLN-pruned. These models are roughly 40% smaller, work only in ComfyUI, and underpin most of the consumer-GPU quants below. For a 24 GB card, `pruned_int8_convrot` (**19.53 GiB**) is the usual starting point. ### Quantized Models MiniMax provides original BF16 checkpoints. The files below are community conversions and repackaged variants, not official MiniMax releases. Confirm compatibility with your runtime's documentation before downloading.

Community FL2VA conversions

| Pruned | Precision | Method | Size | Download | | :---: | :---: | :--- | :---: | :--- | | | ![bf16][badge-bf16] | BF16 | 61.73 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_bf16.safetensors) | | | ![bf16][badge-bf16] | Hybrid (fl2va base + ref2va adaln b15-49) | 20.97 GiB | [![][gh-smhfacct]](https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/resolve/main/minimax_h3_hybrid_fl2va_ref2va_b15-49.safetensors) | | | ![bf16][badge-bf16] | Hybrid (fl2va base + ref2va adaln b20-49) | 20.97 GiB | [![][gh-smhfacct]](https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/resolve/main/minimax_h3_hybrid_fl2va_ref2va_b20-49.safetensors) | | | ![bf16][badge-bf16] | Hybrid (fl2va base + ref2va adaln b25-49) | 20.97 GiB | [![][gh-smhfacct]](https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/resolve/main/minimax_h3_hybrid_fl2va_ref2va_b25-49.safetensors) | | | ![bf16][badge-bf16] | Hybrid (fl2va base + ref2va adaln b30-49) | 20.97 GiB | [![][gh-smhfacct]](https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/resolve/main/minimax_h3_hybrid_fl2va_ref2va_b30-49.safetensors) | | | ![int8][badge-int8] | ConvRot | 31.70 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_int8_convrot.safetensors) | | | ![int8][badge-int8] | ConvRot Lean (HQ) | 21.91 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-HQ.safetensors) | | | ![int8][badge-int8] | ConvRot | 20.94 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot.safetensors) | | | ![int8][badge-int8] | ConvRot Lite | 20.33 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-INT8-ConvRot-Lite.safetensors) | | | ![fp8][badge-fp8] | FP8 E4M3FN | 43.78 GiB | [![][gh-rzgar]](https://huggingface.co/rzgar/minimax_h3_fl2va_fp8_e4m3fn/resolve/main/minimax_h3_fl2va_fp8_e4m3fn.safetensors) | | | ![mxfp8][badge-mxfp8] | MXFP8 | 44.34 GiB | [![][gh-rzgar]](https://huggingface.co/rzgar/minimax_h3_fl2va_fp8_e4m3fn/resolve/main/minimax_h3_fl2va_mxfp8.safetensors) | | | ![fp8][badge-fp8] | FP8 + FP16 attention | 26.70 GiB | [![][gh-rzgar]](https://huggingface.co/rzgar/minimax_h3_fl2va_fp8_e4m3fn/resolve/main/minimax_h3_fl2va_fp16attn_fp8.safetensors) | | | ![nvfp4][badge-nvfp4] | NVFP4 (HQ) | 13.60 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4-HQ.safetensors) | | | ![nvfp4][badge-nvfp4] | NVFP4 | 10.86 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/FL2VA/MiniMax-H3_FL2VA-NVFP4.safetensors) | | | ![nvfp4][badge-nvfp4] | NVFP4 | 32.05 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_fl2va_nvfp4.safetensors) | | | ![int4][badge-int4] | NF4 (DiffSynth) | 15.98 GiB | [![][gh-DiffSynth-Studio]](https://huggingface.co/DiffSynth-Studio/MiniMax-H3-NF4/resolve/main/minimax-h3-fl2va-nf4.safetensors) | | | | OrbitQuant W4A4 | 17.03 GiB | [![][gh-WaveCut]](https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/transformer/diffusion_pytorch_model-00001-of-00005.safetensors) | | | ![int8][badge-int8] | ⚠️ DT-sQKV ConvRot | 21.00 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-DynTime-sQKV/resolve/main/FL2VA/MiniMax-H3_FL2VA-DT-sQKV-INT8-ConvRot.safetensors) | | | ![int8][badge-int8] | ⚠️ DT-sQKV ConvRot Lean | 27.99 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-DynTime-sQKV/resolve/main/FL2VA/MiniMax-H3_FL2VA-DT-sQKV-INT8-ConvRot-HQ.safetensors) | | ✓ | ![bf16][badge-bf16] | BF16 | 37.46 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_bf16.safetensors) | | ✓ | ![fp8][badge-fp8] | FP8 scaled | 19.52 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_fp8_scaled.safetensors) | | ✓ | ![int8][badge-int8] | ConvRot | 19.53 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors) ┊ [![][gh-Abiray]](https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/resolve/main/MiniMax_H3_FL2VA_pruned_int8_convrot.safetensors) | | ✓ | ![nvfp4][badge-nvfp4] | NVFP4 | 18.69 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_fl2va_pruned_nvfp4.safetensors) | | ✓ | ![nvfp4][badge-nvfp4] | NVFP4 + ConvRot INT8 | 18.69 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_fl2va_pruned_nvfp4_convrot_int8.safetensors) | | ✓ | ![nvfp4][badge-nvfp4] | NVFP4 | 11.67 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/resolve/main/MiniMax_H3_FL2VA_pruned_nvfp4.safetensors) | | ✓ | ![int4][badge-int4] | Mixed INT4/INT8 ConvRot | 14.81 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/resolve/main/MiniMax_H3_FL2VA_pruned_mixed_int4_int8_convrot.safetensors) ┊ [![][gh-tsolful]](https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot/resolve/main/minimax_h3_fl2va_pruned_INT4BQ.safetensors) | | ✓ | ![int4][badge-int4] | Mixed INT4/INT8 ConvRot Lean | 17.27 GiB | [![][gh-tsolful]](https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot/resolve/main/minimax_h3_fl2va_pruned_INT4Q.safetensors) | | ✓ | ![int4][badge-int4] | INT4 ConvRot | 15.67 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_fl2va_pruned_int4_convrot_simple.safetensors) | | ✓ | ![int4][badge-int4] | Mixed INT4/INT8 ConvRot | 18.92 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_fl2va_pruned_mixed_int4_int8_convrot_simple.safetensors) | | ✓ | ![int4][badge-int4] | INT4 ConvRot | 16.67 GiB | [![][gh-Merserk]](https://huggingface.co/Merserk/MiniMax-H3-INT4-ConvRot) | | ✓ | ![int4][badge-int4] | INT4 ConvRot (pruned) | 10.56 GiB | [![][gh-Merserk]](https://huggingface.co/Merserk/MiniMax-H3-INT4-ConvRot) | | ✓ | ![int4][badge-int4] | W4A8 ConvRot | 11.68 GiB | [![][gh-AX1Y2JP]](https://huggingface.co/AX1Y2JP/MiniMax-H3-W4A8-ConvRot/resolve/main/minimax_h3_fl2va_pruned_symw4a8convrot.safetensors) ┊ [![][gh-Kijai]](https://huggingface.co/Kijai/MiniMax-H3-experimental/resolve/main/minimax_h3_fl2va_pruned_w4a8_mixed.safetensors) ┊ [![][gh-Winnougan]](https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI/resolve/main/minimax_h3_fl2va_pruned-w4a8_convrot_pruned.safetensors) | *GGUF quants — see the [GGUF section](#gguf).*

Community Ref2VA conversions

| Pruned | Precision | Method | Size | Download | | :---: | :---: | :--- | :---: | :--- | | | ![bf16][badge-bf16] | BF16 | 61.73 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_bf16.safetensors) | | | ![int8][badge-int8] | ConvRot ┊ *(patchin HF 1.02)* | 31.70 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_int8_convrot.safetensors) ┊ [![][gh-t8star]](https://huggingface.co/t8star/minimax_h3_ref2va_patchin_hf102/resolve/main/minimax_h3_ref2va_patchin_hf102_T8.safetensors) | | | ![int8][badge-int8] | ConvRot Lean (HQ) | 21.91 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-HQ.safetensors) | | | ![int8][badge-int8] | ConvRot | 20.94 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot.safetensors) | | | ![int8][badge-int8] | ConvRot Lite | 20.33 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-INT8-ConvRot-Lite.safetensors) | | | ![nvfp4][badge-nvfp4] | NVFP4 (HQ) | 13.60 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4-HQ.safetensors) | | | ![nvfp4][badge-nvfp4] | NVFP4 | 10.86 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-ComfyUI-Quants/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-NVFP4.safetensors) | | | ![nvfp4][badge-nvfp4] | NVFP4 | 32.05 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_ref2va_nvfp4.safetensors) | | | ![nvfp4][badge-nvfp4] | NVFP4 mixed | 22.76 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/resolve/main/MiniMax_H3_Ref2VA_nvfp4_mixed.safetensors) | | | ![int4][badge-int4] | NF4 (DiffSynth) | 15.98 GiB | [![][gh-DiffSynth-Studio]](https://huggingface.co/DiffSynth-Studio/MiniMax-H3-NF4/resolve/main/minimax-h3-ref2va-nf4.safetensors) | | | | OrbitQuant W4A4 | 17.03 GiB | [![][gh-WaveCut]](https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/transformer_ref/diffusion_pytorch_model-00001-of-00005.safetensors) | | | ![nvfp4][badge-nvfp4] | Hybrid NVFP4, FFN-only (Blackwell) | 16.38 GiB | [![][gh-abakanai]](https://huggingface.co/abakanai/Minimax_h3_hybrid/resolve/main/minimax_h3_ref2va_pruned_hybrid_ffn_nvfp4_blackwell.safetensors) | | | ![nvfp4][badge-nvfp4] | Hybrid NVFP4, QKV+FFN (Blackwell) | 14.03 GiB | [![][gh-abakanai]](https://huggingface.co/abakanai/Minimax_h3_hybrid/resolve/main/minimax_h3_ref2va_pruned_hybrid_nvfp4_blackwell.safetensors) | | | ![int8][badge-int8] | ⚠️ DT-sQKV ConvRot | 21.00 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-DynTime-sQKV/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-DT-sQKV-INT8-ConvRot.safetensors) | | | ![int8][badge-int8] | ⚠️ DT-sQKV ConvRot Lean | 27.99 GiB | [![][gh-DmitryDB]](https://huggingface.co/DmitryDB/MiniMax-H3-DynTime-sQKV/resolve/main/Ref2VA/MiniMax-H3_Ref2VA-DT-sQKV-INT8-ConvRot-HQ.safetensors) | | ✓ | ![bf16][badge-bf16] | BF16 | 37.46 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_bf16.safetensors) | | ✓ | ![fp8][badge-fp8] | FP8 scaled | 19.52 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_fp8_scaled.safetensors) | | ✓ | ![int8][badge-int8] | ConvRot | 19.53 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors) ┊ [![][gh-Abiray]](https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/resolve/main/MiniMax_H3_Ref2VA_pruned_int8_convrot.safetensors) | | ✓ | ![nvfp4][badge-nvfp4] | NVFP4 | 18.69 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_ref2va_pruned_nvfp4.safetensors) | | ✓ | ![nvfp4][badge-nvfp4] | NVFP4 + ConvRot INT8 | 18.69 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_ref2va_pruned_nvfp4_convrot_int8.safetensors) | | ✓ | ![nvfp4][badge-nvfp4] | NVFP4 | 11.67 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/resolve/main/MiniMax_H3_Ref2VA_pruned_nvfp4.safetensors) | | ✓ | ![int4][badge-int4] | Mixed INT4/INT8 ConvRot | 14.06 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/Minimax-H3-nvfp4-INT4-INT8-Convrot/resolve/main/MiniMax_H3_Ref2VA_pruned_mixed_int4_int8_convrot.safetensors) ┊ [![][gh-tsolful]](https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot/resolve/main/minimax_h3_ref2va_pruned_INT4BQ.safetensors) | | ✓ | ![int4][badge-int4] | Mixed INT4/INT8 ConvRot Lean | 17.18 GiB | [![][gh-tsolful]](https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot/resolve/main/minimax_h3_ref2va_pruned_INT4Q.safetensors) | | ✓ | ![int4][badge-int4] | INT4 ConvRot | 15.67 GiB | [![][gh-rockerBOO]](https://huggingface.co/rockerBOO/minimax-h3-nvfp4/resolve/main/minimax_h3_ref2va_pruned_int4_convrot_simple.safetensors) | | ✓ | ![int4][badge-int4] | W4A8 ConvRot | 11.68 GiB | [![][gh-AX1Y2JP]](https://huggingface.co/AX1Y2JP/MiniMax-H3-W4A8-ConvRot/resolve/main/minimax_h3_ref2va_pruned_symw4a8convrot.safetensors) ┊ [![][gh-Kijai]](https://huggingface.co/Kijai/MiniMax-H3-experimental/resolve/main/minimax_h3_ref2va_pruned_w4a8_mixed.safetensors) ┊ [![][gh-Winnougan]](https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI/resolve/main/minimax_h3_ref2va_pruned-w4a8_convrot_pruned.safetensors) | *GGUF quants — see the [GGUF section](#gguf).*

Community multi-tier repack

[`DeepBeepMeep/MiniMax-H3`](https://huggingface.co/DeepBeepMeep/MiniMax-H3) collects community files at several precisions. Check the model card for compatibility and licensing.

Quants of the accelerated variants (VDN · PDD · FastH3)

#### GGUF Quantized Models GGUF files are community conversions, not part of the official MiniMax H3 release. Use a compatible loader and follow its documentation.

Community FL2VA GGUF conversions

| Pruned | Quant | Size | Download | | :---: | :---: | :---: | :--- | | | ![Q2_K][badge-Q2_K] | 17.42 GiB † | [![][gh-realrebelai]](https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/resolve/main/MiniMax-H3-FL2VA-Q2_K-(Mixed_Precision).gguf) | | | ![Q3_K_M][badge-Q3_K_M] | 14.50 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q3_K_M.gguf) ┊ [![][gh-realrebelai]](https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/resolve/main/MiniMax-H3-FL2VA-Q3_K_M.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q3_K_M.gguf) | | | ![Q3_K_S][badge-Q3_K_S] | 14.50 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q3_K_S.gguf) | | | ![Q4_0][badge-Q4_0] | 17.36 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q4_0.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q4_0.gguf) | | | ![Q4_1][badge-Q4_1] | 20.41 GiB | [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q4_1.gguf) | | | ![Q4_K_M][badge-Q4_K_M] | 18.50 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q4_K_M.gguf) ┊ [![][gh-realrebelai]](https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/resolve/main/MiniMax-H3-FL2VA-Q4_K_M.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q4_K_M.gguf) | | | ![Q4_K_S][badge-Q4_K_S] | 18.49 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q4_K_S.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q4_K_S.gguf) | | | ![Q5_0][badge-Q5_0] | 21.21 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q5_0.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q5_0.gguf) | | | ![Q5_1][badge-Q5_1] | 24.17 GiB | [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q5_1.gguf) | | | ![Q5_K_M][badge-Q5_K_M] | 22.25 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q5_K_M.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q5_K_M.gguf) | | | ![Q5_K_S][badge-Q5_K_S] | 22.25 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q5_K_S.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q5_K_S.gguf) | | | ![Q6_K][badge-Q6_K] | 26.28 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q6_K.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q6_K.gguf) | | | ![Q8_0][badge-Q8_0] | 33.56 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-FL2VA-Q8_0.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/fl2va/minimax_h3_fl2va-Q8_0.gguf) | | ✓ | ![IQ1_S][badge-IQ1_S] | **3.78 GiB** — smallest DiT published | [![][gh-MarxistLeninist]](https://huggingface.co/MarxistLeninist/MiniMax-H3-FL2VA-Pruned-IQ1-GGUF/resolve/main/minimax_h3_fl2va_pruned-IQ1_S.gguf) | | ✓ | ![IQ1_M][badge-IQ1_M] | 4.22 GiB | [![][gh-MarxistLeninist]](https://huggingface.co/MarxistLeninist/MiniMax-H3-FL2VA-Pruned-IQ1-GGUF/resolve/main/minimax_h3_fl2va_pruned-IQ1_M.gguf) | | ✓ | ![Q2_K][badge-Q2_K] | 6.26 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-Q2_K.gguf) ┊ [![][gh-leejet]](https://huggingface.co/leejet/MiniMax-H3-GGUF) | | ✓ | ![UD-Q2_K_XL][badge-UD-Q2_K_XL] | 7.51 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-UD-Q2_K_XL.gguf) | | ✓ | ![Q3_K_M][badge-Q3_K_M] | 8.16 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-Q3_K.gguf) | | ✓ | ![Q3_K_M][badge-Q3_K_M] | 8.29 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-Pruned-GGUF) | | ✓ | ![UD-Q3_K_XL][badge-UD-Q3_K_XL] | 8.90 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-UD-Q3_K_XL.gguf) | | ✓ | ![Q4_K_M][badge-Q4_K_M] | 10.64 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-Q4_K.gguf) ┊ [![][gh-leejet]](https://huggingface.co/leejet/MiniMax-H3-GGUF) | | ✓ | ![Q5_0][badge-Q5_0] | 12.97 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-Q5_0.gguf) | | ✓ | ![Q6_K][badge-Q6_K] | 15.45 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-Q6_K.gguf) | | ✓ | ![Q8_0][badge-Q8_0] | 19.97 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_fl2va_pruned-Q8_0.gguf) | | ✓ | ![Q8_0][badge-Q8_0] | 20.10 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-Pruned-GGUF) | † `realrebelai` ships Q2_K as a **mixed-precision** build, placing it *above* Q3_K_M. When comparing by size, prioritize the numerical value over the quant name.

Community Ref2VA GGUF conversions

| Pruned | Quant | Size | Download | | :---: | :---: | :---: | :--- | | | ![Q3_K_M][badge-Q3_K_M] | 14.50 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q3_K_M.gguf) ┊ [![][gh-realrebelai]](https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/resolve/main/MiniMax-H3-REF2VA-Q3_K_M.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q3_K_M.gguf) | | | ![Q3_K_S][badge-Q3_K_S] | 14.50 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q3_K_S.gguf) | | | ![Q4_0][badge-Q4_0] | 17.36 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q4_0.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q4_0.gguf) | | | ![Q4_1][badge-Q4_1] | 20.41 GiB | [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q4_1.gguf) | | | ![Q4_K_M][badge-Q4_K_M] | 18.49 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q4_K_M.gguf) ┊ [![][gh-realrebelai]](https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/resolve/main/MiniMax-H3-REF2VA-Q4_K_M.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q4_K_M.gguf) | | | ![Q4_K_S][badge-Q4_K_S] | 18.49 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q4_K_S.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q4_K_S.gguf) | | | ![Q5_0][badge-Q5_0] | 21.21 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q5_0.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q5_0.gguf) | | | ![Q5_1][badge-Q5_1] | 24.17 GiB | [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q5_1.gguf) | | | ![Q5_K_M][badge-Q5_K_M] | 22.25 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q5_K_M.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q5_K_M.gguf) | | | ![Q5_K_S][badge-Q5_K_S] | 22.25 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q5_K_S.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q5_K_S.gguf) | | | ![Q6_K][badge-Q6_K] | 26.28 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q6_K.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q6_K.gguf) | | | ![Q8_0][badge-Q8_0] | 33.56 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/unet/MiniMax-H3-Ref2VA-Q8_0.gguf) ┊ [![][gh-vantagewithai]](https://huggingface.co/vantagewithai/MiniMax-H3-comfyUI-GGUF/resolve/main/ref2va/minimax_h3_ref2va-Q8_0.gguf) | | ✓ | ![Q2_K][badge-Q2_K] | 6.22 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_ref2va_pruned-Q2_K.gguf) | | ✓ | ![Q3_K_M][badge-Q3_K_M] | 8.12 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_ref2va_pruned-Q3_K.gguf) | | ✓ | ![Q4_0][badge-Q4_0] | 10.60 GiB | [![][gh-molbal]](https://huggingface.co/molbal/MiniMax-H3-GGUF) | | ✓ | ![Q4_K_M][badge-Q4_K_M] | 10.60 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_ref2va_pruned-Q4_K.gguf) | | ✓ | ![Q5_0][badge-Q5_0] | 12.94 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_ref2va_pruned-Q5_0.gguf) | | ✓ | ![Q6_K][badge-Q6_K] | 14.00 GiB | [![][gh-molbal]](https://huggingface.co/molbal/MiniMax-H3-GGUF) ‡ | | ✓ | ![Q6_K][badge-Q6_K] | 15.42 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_ref2va_pruned-Q6_K.gguf) | | ✓ | ![Q8_0][badge-Q8_0] | 18.77 GiB | [![][gh-molbal]](https://huggingface.co/molbal/MiniMax-H3-GGUF) § | | ✓ | ![Q8_0][badge-Q8_0] | 19.94 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF/resolve/main/minimax_h3_ref2va_pruned-Q8_0.gguf) ┊ [![][gh-molbal]](https://huggingface.co/molbal/MiniMax-H3-GGUF) | ‡ molbal calls this one **`U16G`** — a hand-tuned mixed layout rather than a stock `Q6_K`. § molbal's **`Q8_CR`** is `Q8_0` with ConvRot applied. Both are in the same repo as the plain quants; read the filename.

Fine-tuned checkpoint quants

`DmitryDB` also publishes stock-compatible quants of community **fine-tunes** of H3 — fine-tuned QKV weights in blocks 0–31 preserved alongside a tested quantization layout, no custom node or core patch required (the ConvRot / NVFP4 tiers match the base-model tiers exactly: 21.91 / 20.94 / 13.60 / 10.86 GiB, plus a ⚠️ DT-sQKV build at 21.00 GiB). The fine-tunes themselves are third-party and are not itemised here; browse [`DmitryDB`'s model list](https://huggingface.co/DmitryDB) directly for the current set.

#### Notes * **`DmitryDB/MiniMax-H3-INT8-Lean-ConvRot`** and **`DmitryDB/MiniMax-H3-ComfyUI-Quants`** are the *same repo* — the author merged and rebranded. Likewise **`…-INT8-Lean-ConvRot-Dynamic-Time-Separate-QKV`** and **`…-DynTime-sQKV`**. Both names in each pair resolve to the same files, so don't download twice. * **`t8star/minimax_h3_ref2va_patchin_hf102`** is a weight *modification*, not a quant: +2 % on the 2×2 spatial high-frequency patch in the video-input projection. The author's own tests showed a weak HF-agent gain and did **not** confirm the "oily/waxy" look was removed. Treat as experimental. * **`Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI`** ships a matching quantized text encoder: [`qwen3vl_32b_minimax_h3-w4a8_convrot.safetensors`](https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI/resolve/main/qwen3vl_32b_minimax_h3-w4a8_convrot.safetensors). * **`unsloth/MiniMax-H3-GGUF`** also carries Qwen3-VL text-encoder GGUFs: `Q2_K_M` 12.2 GiB and `Q4_K_M` 17.0 GiB. * **`DmitryDB/MiniMax-H3-ComfyUI-Quants`** also carries VAE files: video VAE FP16 4.85 GiB, audio VAE FP32 577 MiB. * **VDN and PDD quants are not interchangeable with base quants.** A VDN INT8 file expects the VDN adapters loaded; a SVDQuant PDD build already has the 8-step LoRA folded in, so stacking another acceleration LoRA on top will over-distill the result. * **`DiffSynth-Studio/MiniMax-H3-NF4`** bundles NF4 TE + video VAE + audio VAE. Requires [DiffSynth-Studio](https://github.com/modelscope/DiffSynth-Studio); the project states a **minimum of 8 GB VRAM** on this path. * **`WaveCut/MiniMax-H3-OrbitQuant-W4A4`** bundles a quantized TE and FP32 VAE copies, and requires the [`ComfyUI-OrbitQuant`](https://github.com/iamwavecut/ComfyUI-OrbitQuant/tree/feature/minimax-h3-comfyui) node — the W4A4 path is not loadable without it. [Workflow JSON](https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA.json). ### Text encoders MiniMax-H3 uses **Qwen3-VL-32B** for text and vision. On a 24 GB card, you will usually need to reduce the text encoder after reducing the DiT. The options are listed below. #### Comfy-Org (official repackage) | Model | Precision | Size | Download | | :--- | :---: | :---: | :--- | | `qwen3vl_32b_minimax_h3` | ![bf16][badge-bf16] | 47.97 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_bf16.safetensors) | | `qwen3vl_32b_minimax_h3` | ![int8][badge-int8] | 25.28 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors) | | `qwen3vl_32b_minimax_h3` | ![nvfp4][badge-nvfp4] | **14.61 GiB** | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors) | The `nvfp4_awq` build at **14.61 GiB** is the smallest official TE and the one to pair with a pruned INT8 DiT on a 24 GB card. #### Community quantizations | Model | Precision | Size | Source | | :--- | :---: | :---: | :--- | | `qwen3vl_32b_minimax_h3` | ![Q4_K_M][badge-Q4_K_M] | 13.58 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/text_encoders/qwen3vl_32b_minimax_h3-Q4_K_M.gguf) | | `qwen3vl_32b_minimax_h3` | ![int4][badge-int4] | 13.93 GiB | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_int4_convrot.safetensors) | | `qwen3vl_32b_minimax_h3` | ![nvfp4][badge-nvfp4] | 25.28 GiB † | [![][gh-Abiray]](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors) | | `qwen3vl_32b_minimax_h3` | ![int4][badge-int4] | 15.35 GiB | [![][gh-AX1Y2JP]](https://huggingface.co/AX1Y2JP/MiniMax-H3-W4A8-ConvRot) — W4A8 ConvRot | | `qwen3vl_32b_minimax_h3` | ![int4][badge-int4] | — | [![][gh-Winnougan]](https://huggingface.co/Winnougan/MiniMax-H3-INT4_Convrot_ComfyUI/resolve/main/qwen3vl_32b_minimax_h3-w4a8_convrot.safetensors) — W4A8 ConvRot | | `qwen3vl_32b_minimax_h3` | ![Q2_K][badge-Q2_K] | 12.2 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF) — `Q2_K_M` GGUF | | `qwen3vl_32b_minimax_h3` | ![Q4_K_M][badge-Q4_K_M] | 17.0 GiB | [![][gh-unsloth]](https://huggingface.co/unsloth/MiniMax-H3-GGUF) — `Q4_K_M` GGUF | | `qwen3vl_32b_minimax_h3` | ![Q2_K][badge-Q2_K] | **7.91 GiB** | [![][gh-realrebelai]](https://huggingface.co/realrebelai/MiniMax-H3_GGUFs) — smallest TE published | | `qwen3vl_32b_minimax_h3` | ![int8][badge-int8] | 24.89 GiB | [![][gh-DeepBeepMeep]](https://huggingface.co/DeepBeepMeep/MiniMax-H3) — quanto-INT8 ⚠️ no license | † `Abiray`'s `nvfp4_awq` file is byte-for-byte the size of Comfy-Org's **INT8** build, not of an NVFP4 one. Check the file before assuming it is a smaller download. ### VAE and components #### VAE (video & audio) Both VAEs are **required** for every generation workflow — H3 decodes video and audio through separate autoencoders. | Component | Source | Precision | Size | Download | | :--- | :--- | :---: | :---: | :--- | | Video VAE | Comfy-Org | ![fp16][badge-fp16] | 4.85 GiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors) | | Audio VAE | Comfy-Org | ![fp32][badge-fp32] | 577 MiB | [![][gh-Comfy--Org]](https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors) | | Video VAE | dummy9996 | ![fp8][badge-fp8] | 2.60 GiB | [![][gh-dummy9996]](https://huggingface.co/dummy9996/minimax_h3_vae_fp8/resolve/main/minimax_h3_video_vae_fp8mix.safetensors) | | Audio VAE | dummy9996 | ![bf16][badge-bf16] | 289 MiB | [![][gh-dummy9996]](https://huggingface.co/dummy9996/minimax_h3_vae_fp8/resolve/main/minimax_h3_audio_vae_bf16.safetensors) | | Video VAE | Kijai | ![int8][badge-int8] | 2.95 GiB | [![][gh-Kijai]](https://huggingface.co/Kijai/MiniMax-H3-experimental/resolve/main/minimax_h3_video_vae_int8_convrot.safetensors) | The `fp8mix` video VAE (**2.60 GiB**) plus the `bf16` audio VAE (**289 MiB**) save roughly 2.5 GiB over the official pair — worth taking on a 12–16 GB card, where the VAE competes with the DiT for the same headroom. #### Tiny Autoencoder (TAE) — previews only A quickly-trained 2D tiny VAE by [Kijai](https://huggingface.co/Kijai/MiniMax-H3-TAE). The author's own assessment: not a great outcome, but it still beats `latent2rgb` for previews. **9 MiB.** Currently only usable through the `ModelPreviewOverride` node in [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes). | Component | Size | Download | | :--- | :---: | :--- | | TAE (preview VAE) | 9 MiB | [![][gh-Kijai]](https://huggingface.co/Kijai/MiniMax-H3-TAE/resolve/main/vae_approx/taeh3.safetensors) | #### Image VAE (single-frame) An experimental image-specialised H3 VAE that decodes a single temporal latent (`T=1`) into one still. Merged checkpoint — no custom node needed. ⚠️ **For image workflows only.** The image-tuned decoder materially regresses multi-frame video reconstruction, so keep the original VAE loaded for video. | Component | Size | Download | | :--- | :---: | :--- | | Single-image VAE (step 1597) | 4.85 GiB | [![][gh-Mamad8]](https://huggingface.co/Mamad8/MiniMax-H3-Image-VAE/resolve/main/minimax_h3_t1_image_vae_step1597.safetensors) | #### Ref Patch — FL2VA that behaves more like Ref2VA Diffs the **112 keys shared** between the `ref2va` and `fl2va` weights and stores the differences as a single 148 MiB patch, letting the lighter FL2VA checkpoint partially mimic Ref2VA behaviour. Apache-2.0. Requires the [`ComfyUI-MiniMaxH3_Ref-Patch`](https://github.com/lihaoyun6/ComfyUI-MiniMaxH3_Ref-Patch) node. | Component | Size | Download | | :--- | :---: | :--- | | Ref Patch | 148 MiB | [![][gh-lihaoyun6]](https://huggingface.co/lihaoyun6/MiniMax-H3-Ref-Patch) | #### Fun Control — Canny / Depth / HED / MLSD / Pose + inpainting [`alibaba-pai/MiniMax-H3-Fun-Controlnet-Union`](https://huggingface.co/alibaba-pai/MiniMax-H3-Fun-Controlnet-Union) adds structural control to H3 without retraining the base model. It is a **single ~6.8 GiB control branch** — one `control_proj_in` plus five `control_blocks` grafted onto layers 0, 10, 20, 30 and 40 of the DiT's 50 — and one checkpoint covers every control type rather than one file per modality. | | | | :--- | :--- | | Control types | Canny · Depth · HED · MLSD · Pose, plus video inpainting and outpainting | | Control input | `control_in_dim = 49` — the control video and the inpaint mask enter through the same projection | | Strength | `control_context_scale` from **0** (ignore) to **1** (follow closely) | | Guidance | **Guidance-distilled**: run at `guidance_scale = 1.0`. Raising CFG does not help and costs a second forward pass. | | Frame count | Snaps to `17*n + 5`, up to 15 s at 24 fps | | Upstream | [`aigc-apps/VideoX-Fun`](https://github.com/aigc-apps/VideoX-Fun) (★2244, Apache-2.0) — the training and inference code | | ComfyUI | [`wyzborrero/ComfyUI-H3-FunControl`](https://github.com/wyzborrero/ComfyUI-H3-FunControl) (★34, Apache-2.0). Core ComfyUI support is open as draft PR [#15860](https://github.com/comfyanonymous/ComfyUI/pull/15860). | ### Runtimes | Engine | ⭐ | H3 support | | :--- | ---: | :--- | | [`ComfyUI`](https://github.com/comfyanonymous/ComfyUI) | 127159 | Native, day-0. INT8 is now in mainline (commit `1a510f04`) — see [Compatibility](#compat) before reusing older INT8 quants. | | [`modelscope/DiffSynth-Studio`](https://github.com/modelscope/DiffSynth-Studio) | 12925 | `MiniMaxH3Pipeline` in `diffsynth.pipelines.minimax_h3_audio_video`; docs at `docs/en/Model_Details/MiniMax-H3.md`, examples at `examples/minimax_h3/`. Ships **NF4** quantized inference with an **8 GB VRAM** floor. | | [`ModelTC/LightX2V`](https://github.com/ModelTC/LightX2V) | 2655 | Full inference support: parallelism, quantized DiT, feature caching. Scripts at `scripts/minimax_h3`. Also the home of the Turbo 4-step / 768p LoRAs. | | [`NVlabs/Sana` Sol-Engine](https://github.com/NVlabs/Sana/tree/sol-engine/models/minimax_h3) | 8812 | Hardware-specific full-stack runtimes (kernel optimization, Sol-Attn, approximate caching, memory-efficient decoding) on the `sol-engine` branch, **Apache-2.0**. Released full-opt configs for GB200, GB10, H100, A100, RTX 4090 and RTX 5090 with **3.55–4.52x** speedup vs the matching baseline on the same hardware (e.g. 4.44x on a single RTX 4090, 768p@5s); launch via `scripts/launch_config.py`, local or Slurm. The branch now also ships **Sol-H3**, a standalone datacentre runtime — see [Sol-H3](#solh3). | | [`OpenVDN/vdn-minimax-h3`](https://github.com/OpenVDN/vdn-minimax-h3) | 397 | **Video Delta Net** — a hybrid-attention rebuild of H3 that runs a frame-wise linear-attention branch alongside the softmax branch, shipped as plug-and-play LoRA adapters merged at inference. Apache-2.0, with training *and* inference code open, not just weights. See [VDN](#vdn). | | [`MiniMax-AI/MiniMax-H3`](https://github.com/MiniMax-AI/MiniMax-H3) | 5536 | The official repository — reference implementation and prompt guides. | | [`antirez/h3.c`](https://github.com/antirez/h3.c) | 1652 | Apple Silicon native Metal engine, **MIT**, tutorial in the README. T2V/A, first-last-frame, and ordered Ref2VA references all working. | | [`MiniMaxH3ComfyUI/MiniMax-H3-ComfyUI`](https://github.com/MiniMaxH3ComfyUI/MiniMax-H3-ComfyUI) | 101 | Runs the 33B + Turbo LoRA locally with SGLang / vLLM / diffusers as selectable backends; T2V / I2V / R2V templates included. | | [`unslothai/unsloth`](https://github.com/unslothai/unsloth) | 70709 | Lists MiniMax-H3 among the models it can run and train. | ## ComfyUI workflows & nodes ### Conditioning & orchestration | Node | ⭐ | What it does | | :--- | ---: | :--- | | [`comfyui-minimax-h3-audio-T8`](https://github.com/T8mars/comfyui-minimax-h3-audio-T8) ![Conditioning][cat-cond] | 653 | v1.17.0, **62 nodes** across eight menus: Audio (stable), Audio Experimental (multi-rate), Still, Conditioning, Models, Long Video, Speech, Source AV. Baseline ComfyUI `0.31.0`, commit `cbbc9dab1`, Python 3.10+. | | [`ComfyUI_MiniMaxH3_Director`](https://github.com/huangserva/ComfyUI_MiniMaxH3_Director) ![Conditioning][cat-cond] | 553 | Provides five importable JSON templates: t2v, fl2v, r2v, v2v, and rv2v. | | [`ComfyUI-H3-Motion-Context`](https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context) ![Conditioning][cat-cond] | 491 | Chains clips to maintain motion and sound continuity: clip A's last frames and audio inform clip B's start. **Patches at runtime only**; re-validates against current ComfyUI source on each start, refusing to run on mismatch — the safest patching approach in this list. | | [`ComfyUI_MiniMaxH3_Director`](https://github.com/AIMixer/ComfyUI_MiniMaxH3_Director) ![Conditioning][cat-cond] | 359 | The original Director. | | [`ComfyUI-MiniMaxH3-Easy`](https://github.com/nkxx188/ComfyUI-MiniMaxH3-Easy) ![Conditioning][cat-cond] | 332 | Provides a compact workflow for T2V, I2V, first/last-frame, and reference video. Supports unified multi-media input with `@` references and inline dialogue blocks. | | [`ComfyUI-MiniMaxH3-Director`](https://github.com/seesee75-commits/ComfyUI-MiniMaxH3-Director) ![Conditioning][cat-cond] | 182 | Offers a timeline editor: drag media onto tracks, trim on a ruler, assign one prompt per shot, with live sampling preview, retakes, and shot chaining. The compiled final prompt remains visible during editing. | | [`ComfyUI-PainterNodes`](https://github.com/princepainter/ComfyUI-PainterNodes) ![Conditioning][cat-cond] | 178 | `MiniMaxRefToVideo2` node supports the official reference and dialogue format. | | [`ComfyUI-MiniMaxDirector`](https://github.com/imbutus/ComfyUI-MiniMaxDirector) ![Conditioning][cat-cond] | 49 | Timeline editor: shots, camera moves and audio cues on a track, compiled into the single structured prompt from the official guides — camera as motion type × amplitude × speed, voiceovers, dialogue that crosses a cut, on-screen text — plus a WHO & WHAT card per named thing (face, voice, retention, face swap) and a linter for the rules the guides state. MIT. | | [`ComfyUI-H3-FunControl`](https://github.com/wyzborrero/ComfyUI-H3-FunControl) ![Conditioning][cat-cond] | 34 | Loads [`MiniMax-H3-Fun-Controlnet-Union`](#funcontrol) and exposes `control_context_scale` plus the Canny / Depth / HED / MLSD / Pose preprocessors and the inpaint mask path. Apache-2.0. Core ComfyUI support is still a draft PR, so this is the working route today. | | [`OpenH3-IR`](https://github.com/ruashots/open-h3-ir) ![Conditioning][cat-cond] | 19 | The ComfyUI side of OpenH3-IR (see [Prompting](#recipes-prompt)), using its Context-IR service from the same repo, on a model you already run. Drop your pictures, clips and sounds on one panel, name them, then mention them by name in the sentence. The nodes and the service are one tool listed twice, not two projects. | ### Upscaling, loading & repair | Node | ⭐ | What it does | | :--- | ---: | :--- | | [`scraed/LanPaint`](https://github.com/scraed/LanPaint) ![Conditioning][cat-cond] | 1331 | Performs training-free video and audio inpainting; H3 support was fixed in v2.1.0. | | [`ComfyUI-MiniMaxH3_LatentUpscaler`](https://github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler) ![Upscaling][cat-upscale] | 191 | Latent spatial upscaler for H3's `NestedTensor` AV latents (video `[B,24,T,H/16,W/16]` + audio `[B,32,2,T_audio]`), which stock `LatentUpscaleBy` cannot process. Re-noises video and audio for two-pass sampling and scales `minimax_refs` / `minimax_keyframes` conditioning. `audio_denoise`: **0** locks audio, **1** fully remixes, **0.25–0.5** for light remixing. | | [`Comfyui_Minimax_h3_latent_Upscaler`](https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler) ![Upscaling][cat-upscale] | 151 | **Learned** neural upscaler for H3's 24-channel latents — generate at low resolution, upscale the latent in place, then refine at target resolution, skipping the `decode → pixel upscale → encode` round-trip through the ~5B VAE and the ghosting that naive latent interpolation introduces. Continuous **1.0–4.0×** scale. [Checkpoints on HF](https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler) (bf16/fp16 ~691 MB, fp32 ~1.38 GB, Apache-2.0) with I2V and R2V example workflows; trained on ~80k paired samples. | | [`ComfyUI-H3VAE_TRT`](https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT) ![Acceleration][cat-accel] | 212 | ONNX / TensorRT replacement for H3's ~5B video VAE — the decode stage, not the DiT. The author reports **up to 1.7×** on VAE decode. Apache-2.0, released 2026-09-01. Engines are built per GPU on first run. | | [`ComfyUI-INT8-Fast`](https://github.com/BobJohnson24/ComfyUI-INT8-Fast) ![Acceleration][cat-accel] | 286 | **Largely superseded** as INT8 is now native in ComfyUI. Its remaining value is `convert_comfy_quant.py`; see [Compatibility](#compat). | ### Templates & example workflows Official ComfyUI templates (these ship with ComfyUI; links allow viewing the graph without launching the app): * [Text-to-Video (T2V)](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_t2v.json) · [Image-to-Video (I2V)](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_i2v.json) · [Reference-to-Video (R2V)](https://github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json) Community workflows: * [MiniMax-H3 FL2V GGUF workflow](https://huggingface.co/Abiray/MiniMax-H3-GGUF/resolve/main/minimax_fl2v_gguf_workflow.json) — Loads and runs the GGUF-quantized FL2VA model. * [`joeygambino/MiniMax-H3-Multishot-Workflow`](https://huggingface.co/joeygambino/MiniMax-H3-Multishot-Workflow) — Strings several FL2VA/Ref2VA clips into one continuous sequence with matched audio handoffs. Apache-2.0 licensed. * [`javawock7618/comfy-MiniMax-H3-workflows`](https://huggingface.co/javawock7618/comfy-MiniMax-H3-workflows) — Bundles the entire low-VRAM acceleration stack into one importable workflow. * [OrbitQuant T2VA](https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA.json) · [T2VA API form](https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/workflows/MiniMax-H3-OrbitQuant-T2VA-api.json) · [Ref2VA API form](https://huggingface.co/WaveCut/MiniMax-H3-OrbitQuant-W4A4/resolve/main/comfyui/workflows/MiniMax-H3-OrbitQuant-Ref2VA-api.json) — require [`ComfyUI-OrbitQuant`](https://github.com/WaveCut/ComfyUI-OrbitQuant). ## Prompting H3 prompts have a fixed three-part structure, inline `