Files read on Hugging Face: 2026-10-24

Wan2.2 TI2V 5B on Mac: files, RAM and published times

Wan2.2 TI2V 5B is a dense 5B model for text-to-video and image-to-video at 720P and 24 fps. On a Mac it runs through mlx-video, and FastVideo publishes a distilled INT8 derivative.

Wan-AI · Updated 2026-10-24
Answer

Wan2.2 TI2V 5B runs on a Mac with 24 GB or more. A published run on M5 · 24 GB used the FastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model) build (source, 2026-08-10).

Smallest Mac with a published run
24 GBM5 · 24 GB · FastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)
Smallest complete download
13.55 GBQuantStack GGUF Q8_0
Files read
2026-10-24Hugging Face

Apache License 2.0. Commercial use: allowed. Excluded territories: none listed.

See the exact GGUF files and sizes

Table 1

What do you download?

Only builds with every required component and an exact probed file size are shown.

Comfy-Org repackaged FP16 with FP8 scaled UMT5

safetensors · FP16 · 18.14 GB complete · Recommended

Mac routes: ComfyUI

Exact component files for Comfy-Org repackaged FP16 with FP8 scaled UMT5
FileComponentRequiredDownloadSource
split_files/diffusion_models/wan2.2_ti2v_5B_fp16.safetensorsComfy-Org/Wan-2.2TI2V-5B dense transformerditYes10.00 GBHugging FaceRead 2026-10-24
split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensorsComfy-Org/Wan-2.2UMT5-XXL text encodertext_encoderYes6.74 GBHugging FaceRead 2026-10-24
split_files/vae/wan2.2_vae.safetensorsComfy-Org/Wan-2.2Wan2.2-VAEvaeYes1.41 GBHugging FaceRead 2026-10-24

Build sources: Comfy-Org Wan 2.2 repackaged card.

QuantStack GGUF Q8_0

gguf · Q8_0 · 13.55 GB complete

Mac routes: ComfyUI

Exact component files for QuantStack GGUF Q8_0
FileComponentRequiredDownloadSource
Wan2.2-TI2V-5B-Q8_0.ggufQuantStack/Wan2.2-TI2V-5B-GGUFTI2V-5B dense transformerditYes5.40 GBHugging FaceRead 2026-10-24
split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensorsComfy-Org/Wan-2.2UMT5-XXL text encodertext_encoderYes6.74 GBHugging FaceRead 2026-10-24
split_files/vae/wan2.2_vae.safetensorsComfy-Org/Wan-2.2Wan2.2-VAEvaeYes1.41 GBHugging FaceRead 2026-10-24

Build sources: QuantStack Wan2.2 TI2V 5B GGUF card · Comfy-Org Wan 2.2 repackaged card.

FastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)

mlx · INT8 3-step DMD2 · 19.50 GB complete

Mac routes: FastVideo MLX (FastMetal)

Exact component files for FastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)
FileComponentRequiredDownloadSource
mlx_dit.safetensorsFastVideo/FastMetal-5B-QADTI2V-5B dense transformerditYes5.31 GBHugging FaceRead 2026-10-24
text_encoder/model-00001-of-00003.safetensorsFastVideo/FastMetal-5B-QADUMT5-XXL text encodertext_encoderYes4.94 GBHugging FaceRead 2026-10-24
text_encoder/model-00002-of-00003.safetensorsFastVideo/FastMetal-5B-QADUMT5-XXL text encodertext_encoderYes4.98 GBHugging FaceRead 2026-10-24
text_encoder/model-00003-of-00003.safetensorsFastVideo/FastMetal-5B-QADUMT5-XXL text encodertext_encoderYes1.44 GBHugging FaceRead 2026-10-24
vae/diffusion_pytorch_model.safetensorsFastVideo/FastMetal-5B-QADWan2.2-VAEvaeYes2.82 GBHugging FaceRead 2026-10-24

Build sources: FastMetal-5B-QAD card.

Official Wan-AI weights (unquantized)

safetensors · unquantized · 34.18 GB complete

Mac routes: Blaizzy/mlx-video

Exact component files for Official Wan-AI weights (unquantized)
FileComponentRequiredDownloadSource
diffusion_pytorch_model-00001-of-00003.safetensorsWan-AI/Wan2.2-TI2V-5BTI2V-5B dense transformerditYes9.83 GBHugging FaceRead 2026-10-24
diffusion_pytorch_model-00002-of-00003.safetensorsWan-AI/Wan2.2-TI2V-5BTI2V-5B dense transformerditYes10.00 GBHugging FaceRead 2026-10-24
diffusion_pytorch_model-00003-of-00003.safetensorsWan-AI/Wan2.2-TI2V-5BTI2V-5B dense transformerditYes0.18 GBHugging FaceRead 2026-10-24
models_t5_umt5-xxl-enc-bf16.pthWan-AI/Wan2.2-TI2V-5BUMT5-XXL text encodertext_encoderYes11.36 GBHugging FaceRead 2026-10-24
Wan2.2_VAE.pthWan-AI/Wan2.2-TI2V-5BWan2.2-VAEvaeYes2.82 GBHugging FaceRead 2026-10-24

Build sources: Wan2.2 TI2V 5B model card · mlx-video Wan README.

Table 2

Does it fit your Mac's RAM?

Published names the exact measured machine. est. is a memory calculation. unknown means the run verdict is not established, not that the model fails.

Run verdicts by Mac memory tier
Mac RAMComplete buildDownloadRun verdictBasis
8 GBComfy-Org repackaged FP16 with FP8 scaled UMT5FP16 · ComfyUI18.14 GBDoes not fitest.
16 GBQuantStack GGUF Q8_0Q8_0 · ComfyUI13.55 GBunknownunknown
24 GBFastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)INT8 3-step DMD2 · FastVideo MLX (FastMetal)19.50 GBRuns · PublishedPublished13-inch MacBook Air · Apple M5 · 24 GB RAM
32 GBFastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)INT8 3-step DMD2 · FastVideo MLX (FastMetal)19.50 GBRuns · Published on M5 · 24 GBPublished13-inch MacBook Air · Apple M5 · 24 GB RAM
48 GBFastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)INT8 3-step DMD2 · FastVideo MLX (FastMetal)19.50 GBRuns · Published on M4 Max · 36 GBPublishedApple M4 Max · 36 GB RAM
64 GBFastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)INT8 3-step DMD2 · FastVideo MLX (FastMetal)19.50 GBRuns · Published on M4 Max · 36 GBPublishedApple M4 Max · 36 GB RAM
96 GBFastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)INT8 3-step DMD2 · FastVideo MLX (FastMetal)19.50 GBRuns · Published on M4 Max · 36 GBPublishedApple M4 Max · 36 GB RAM
128 GBFastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model)INT8 3-step DMD2 · FastVideo MLX (FastMetal)19.50 GBRuns · Published on M4 Max · 36 GBPublishedApple M4 Max · 36 GB RAM

How does it rank in blind votes?

No independent score yet.

Table 3

What published Mac times are available?

Third-party results keep their original machine and settings. ModelFit does not run these benchmarks.

Published Mac generation measurements
Machine and RAMSettingsAppResultSource
MacBook Pro · Apple M3 Max · 64 GB RAM1280x704 · 5 s · 40 steps · 121 frames · 24 fps · unquantized · I2V with an input image, 121 frames, 5 seconds as the author describes it. The run date is not stated; the date shown is the article date.Blaizzy/mlx-videoSuccess96 minutes57 GB peakrotecodefraktion: local Wan video on a MacPublished 2026-10-10 · Grade A
MacBook Pro · Apple M3 Max · 64 GB RAM1280x704 · 1.7 s · 40 steps · 41 frames · 24 fps · unquantized · T2V, 41 frames. Time is the article total: encoder load, transformer load, 40 diffusion steps and VAE decode. The run date is not stated; the date shown is the article date.Blaizzy/mlx-videoSuccess17:4356.4 GB peakrotecodefraktion: local Wan video on a MacPublished 2026-10-10 · Grade A
Apple M4 Max · 36 GB RAM1280x704 · 3 steps · 81 frames · INT8 3-step DMD2 · FastMetal-QAD derivative, INT8 DiT, TAEHV decode. End-to-end time includes prompt encoding, DiT load, denoising and decode/export. The date is the article date; the run date is not stated.FastVideo MLX (FastMetal)Success151.4 s9.34 GiBFastMetal-QAD announcementPublished 2026-08-10 · Grade A
13-inch MacBook Air · Apple M5 · 24 GB RAM1280x704 · 3 steps · 81 frames · INT8 3-step DMD2 · FastMetal-QAD derivative, Baseline mode, average over repeated runs. The date is the article date; the run dates are not stated.FastVideo MLX (FastMetal)Success200.1 s9.54 GiBFastMetal-QAD announcementPublished 2026-08-10 · Grade A

Which Mac software routes are listed?

  • ComfyUI

    Official

    Formats: safetensors, gguf. Memory policy: component stages stay resident.

    On Apple Silicon ComfyUI runs on PyTorch MPS, not MLX. GGUF files need the separate city96/ComfyUI-GGUF custom node. Resident size is the sum of the weights, an upper bound: ComfyUI can unload models between stages and publishes no per-model Mac peak.

    Sources: ComfyUI README · ComfyUI system requirements · ComfyUI CLI arguments · ComfyUI-GGUF README

  • FastVideo MLX (FastMetal)

    Community

    Formats: mlx. Memory policy: component stages stay resident.

    FastMetal checkpoints are distilled INT8 derivatives; their times are not base-model figures.

    Sources: FastVideo MLX install guide

  • Blaizzy/mlx-video

    Community

    Formats: mlx, safetensors. Memory policy: component stages stay resident.

    Sources: mlx-video Wan README

What does the official configuration say?

"The official native runner needs a GPU with at least 24GB VRAM. No official Mac memory minimum was found."

Source: Wan2.2 TI2V 5B model card · Read 2026-10-24.

What known issues should you check?

  • Known issue 1

    FastMetal 5B QAD is a three-step distilled INT8 derivative, so its times and memory are not base-model figures.

    Source: FastMetal-QAD announcement · Read 2026-10-24

  • Known issue 2

    mlx-video lists 40 steps for TI2V-5B while the official Wan config uses 50, so pass the step count explicitly when comparing times.

    Source: Wan2.2 TI2V 5B config · Read 2026-10-24

Frequently asked questions

How fast is Wan2.2 5B on a Mac?

The full model through mlx-video took 17:43 for a 41-frame 1280x704 clip at 40 steps on an M3 Max with 64 GB, peaking at 56.4 GB. The distilled FastMetal 5B derivative took 151.42 s for an 81-frame clip on an M4 Max with 36 GB, so the two are not comparable.

What is the FastMetal 5B build?

FastMetal 5B QAD is a DMD2-distilled, 3-step, INT8 derivative of Wan2.2 TI2V 5B published by FastVideo. It is not the base 50-step model, so its times and memory describe the derivative only.

Sources