Wan 2.2 on Mac: Which Model Fits, Files and Published Timings (2026)

On a Mac, Wan 2.2 TI2V 5B is the model to start with. The full-weights 5B text-to-video run peaked at 56.4 GB on a 64 GB MacBook Pro (rotecodefraktion, 2026), while FastVideo's three-step INT8 FastMetal derivative peaked at 9.34 GiB on an M4 Max (Hao AI Lab, 2026). A14B has one published Mac benchmark: 10:25 for the Q4 pack at 81 frames and 512×320 on an M5 Max with 128 GB (PocketAI, 2026).

ModelFit runs no video benchmarks. Every time and peak below belongs to the source named beside it, with its own machine and settings, and none is averaged across settings. The sources were read on 2026-10-24. The Wan 2.2 5B and Wan 2.2 A14B model pages list each complete build with its RAM verdict. For other models see the video hub, and for the app-by-app picture see local video generation on a Mac.

Which Wan 2.2 model fits a Mac?

TI2V 5B is the Mac starting point. It is one dense 5B model that takes text or an image, and every Mac run below used it except PocketAI's A14B packages. A14B is two experts of about 14B parameters each, 27B in total with only 14B active per step (Wan README, 2026). The name does not describe a 14B download.

The high-noise expert runs the early steps and sets the overall layout. The low-noise expert runs the later steps and refines detail. A Mac route therefore has to carry both expert files.

Frames, fps and steps come from the repository config files (shared, A14B, 5B). The VRAM notes come from the model cards (5B, A14B) and describe the native PyTorch runner on a GPU. They say nothing about Mac unified memory.

Wan's README also says the 5B makes a 5-second 720P video in under 9 minutes on "a single consumer-grade GPU". The GPU is not named, so that figure says nothing about any Mac.

What do the Wan 2.2 files weigh?

The official A14B text-to-video repository holds 126,177,685,050 bytes of weights: two experts of about 57 GB each, an 11 GB UMT5 text encoder and a 0.51 GB VAE (file listing, 2026). That is a download size. The card's 80GB figure is a GPU requirement, and most Mac setups start from smaller files.

Sizes below are decimal GB from the Hugging Face file listings, rounded, and every A14B row needs both experts. The ComfyUI text encoder and VAE are umt5_xxl_fp8_e4m3fn_scaled.safetensors and wan_2.1_vae.safetensors.

The official experts are twice the size of ComfyUI's fp16 files. mlx-video's converter still reads the official PyTorch checkpoints, so that route starts with the full download, while PocketAI's package skips it (mlx-video, 2026). PocketAI states its model data totals 26.68 GiB (PocketAI card, 2026).

The 5B set is much smaller. The ComfyUI files come from Comfy-Org and the others from the repositories named in the first column.

Which GGUF files does Wan 2.2 need?

A14B needs two GGUF files of the same quant: one in HighNoise/ and one in LowNoise/. Q4_K_M is 9,650,090,496 bytes per file, 19 GB for the pair. QuantStack's T2V card calls the files a direct conversion of the base model and says to place them in ComfyUI/models/unet (QuantStack card, 2025). ComfyUI-GGUF swaps only the diffusion model loader, so the text encoder and VAE stay separate.

The T2V names follow HighNoise/Wan2.2-T2V-A14B-HighNoise-<quant>.gguf and LowNoise/Wan2.2-T2V-A14B-LowNoise-<quant>.gguf, and the two files of a quant have identical sizes (file listing, 2026). The I2V repository uses the same pattern with I2V in the name, and each quant below is 1,638,400 bytes larger there.

The 5B has no expert split. Wan2.2-TI2V-5B-Q4_K_M.gguf is 3,433,116,000 bytes and Wan2.2-TI2V-5B-Q8_0.gguf is 5,400,179,040 bytes (5B listing, 2026). Loading them needs the ComfyUI-GGUF node pack, which the next section covers.

Which software runs Wan 2.2 on a Mac?

Seven routes have some Mac evidence. mlx-video, PocketAI's prebuilt MLX packages and FastVideo's FastMetal come with timings, while ComfyUI has only issue reports. The table lists what each route publishes.

mlx-video

mlx-video converts the official weights to MLX and can quantize the transformer to 4 or 8 bit with --quantize --bits 4 --group-size 64. Its 5B config defaults to 40 steps while Wan's native config uses 50, frame counts must be 4n+1 and the resolution must divide by 32, so pass every setting explicitly (mlx-video README, 2026). The module path is mlx_video.models.wan_2; the README still shows older names in places. These mirror the commands behind the M3 Max timing below, with a shorter prompt:

uv venv --python 3.12
uv pip install "mlx-video @ git+https://github.com/Blaizzy/mlx-video.git" torch
uvx --from huggingface_hub hf download Wan-AI/Wan2.2-TI2V-5B --local-dir ~/models/Wan2.2-TI2V-5B
.venv/bin/python -m mlx_video.models.wan_2.convert --checkpoint-dir ~/models/Wan2.2-TI2V-5B --output-dir ~/models/Wan2.2-TI2V-5B-MLX
.venv/bin/python -m mlx_video.models.wan_2.generate --model-dir ~/models/Wan2.2-TI2V-5B-MLX \
  --prompt "A chestnut horse runs across a sunlit field" \
  --width 1280 --height 704 --num-frames 41 --steps 40 --seed 42 --output-path clip.mp4

PocketAI MLX packages

PocketAI publishes A14B T2V as ready-to-run MLX packages: Q4 on the main revision, plus q8 and bf16 revisions. It describes itself as an unofficial community derivative. Its card lists 24 GB of unified memory for short low-guidance tests and says 128 GB was used for the validated 81-frame benchmark (PocketAI card, 2026). It needs Python 3.11 to 3.13 and ffmpeg, and its example pocketai-wan-video command uses --steps 10 --guide-scale 3,4 --scheduler unipc.

FastMetal

FastMetal is a different model from the base 5B. It is a three-step student trained with DMD2 and quantization-aware training on an INT8 grid, so its speed and quality do not describe Wan's 50-step default (Hao AI Lab, 2026). FastVideo lists the 5B at 16 GB of unified memory and up, on macOS 14 or newer with Python 3.12 (FastVideo docs, 2026). Its FastMetal 14B is built from Wan2.1-T2V-14B, so none of its figures describe A14B.

ComfyUI and ComfyUI-GGUF

ComfyUI is the least settled route. The checked-in 5B workflow loads wan2.2_ti2v_5B_fp16.safetensors and uses 20 sampler steps with model shift 8, where Wan's native config uses 50 steps and shift 5.0 (workflow, 2026). The A14B template loads fp8_scaled experts (template, 2026), and three GitHub issues describe what happens on Apple Silicon:

• Issue 9255, closed 2025-08-14: the "Wan 2.2 14B Text to Video" template failed on an M2 Mac with "Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype". Two commenters said the QuantStack GGUF files worked instead, one naming Q5_K_M, and neither gave a timing.
• Issue 15010, filed 2026-07-20 and still open when checked: the native 5B image-to-video workflow finished on an M4 Max with 64 GB, but a 49-frame, 512×896, 16 fps render developed vertical bands and oversaturation from about frame 16.
• Issue 14837, open: PyTorch's MPS backend silently corrupts output when one attention matrix reaches about 2^31 elements, which long or high-resolution video can hit.

To use GGUF files, clone ComfyUI-GGUF into ComfyUI/custom_nodes, run pip install --upgrade gguf, place the .gguf files in ComfyUI/models/unet and use its GGUF Unet loader. The A14B template wires one loader per expert, so a GGUF swap replaces both. The ComfyUI on Mac guide covers the install.

What have people measured on a Mac?

Four sources publish Wan 2.2 timings with a machine and settings: PocketAI for A14B, FastVideo and an independent M3 Max write-up for the 5B, and a Rapid-MLX benchmark section with thinner detail. None of the A14B figures carries a run date.

PocketAI's README reports an Apple M5 Max with 128 GB, identical prompt and seed, 512×320, 81 frames, 16 fps, 10 UniPC steps and guidance 3/4. Its page carries no run date and was last modified 2026-08-15. Ten steps is a quarter of A14B's native 40, and 512×320 is below the 480P and 720P sizes the card lists. The three precisions finish within 52 seconds of each other, while peak memory spans 16.37 to 41.48 GiB. PocketAI calls Q4 the recommended default.

The 5B rows come from three sources. The first four are FastVideo's FastMetal post of 2026-08-10: 1280×704, 81 frames, the three-step INT8 derivative, an M4 Max and a 13-inch MacBook Air with an M5 and a 10-core GPU. The next two are the independent write-up of 2026-10-10 on a MacBook Pro with an M3 Max (40 GPU cores), macOS 27.2, MLX 0.32.3 and mlx-video 0.0.1, running the full model at 40 steps. The last row is the Rapid-MLX README, undated, which states neither quantization nor steps.

These rows are not a chip comparison. The FastMetal runs use a distilled model, and the M4 Max baseline includes a cold prompt encode of about 47 seconds, while the post says fast-mode rows reuse a cached prompt embedding and generate every Nth frame before interpolating the rest (Hao AI Lab, 2026). The M3 Max rows run the full model: its text-to-video run spent 15:06 in diffusion at 22.6 seconds per step and 2:01 decoding. The author wrote that a Mac with 64 GB is enough and one with 32 GB is not (rotecodefraktion, 2026).

The same article ran PyTorch diffusers on the same Mac. Diffusion took 28.6 seconds per step against 22.6 for mlx-video, then macOS killed the process while decoding, with demand reaching 75 GB, or 95 GB with tiling enabled. The article also projects M4 Max, M5 Max and M6 Max times. Only the M3 Max was measured, so the projections are left out here.

What is still unknown about Wan 2.2 on a Mac?

Four gaps matter for a buying or download decision:

• Wan gives no Mac memory minimum. The 80GB and 24GB notes are GPU runner statements, and neither the README nor the model cards mention a Mac figure.
• PocketAI's M5 Max table has no measurement date, and no source times A14B at 480P or 720P on a Mac.
• No A14B timing through ComfyUI-GGUF exists in what was found, only two comments in one issue saying GGUF worked, with Q5_K_M named. mlx-video documents a 4-step Lightning LoRA path for A14B, but no Mac timing for it was found.
• PocketAI's Q8 and BF16 quality notes rest on one prompt and one seed, and none of the sources read here compares GGUF quants for quality.

For hardware advice across the video models, see the best Mac for AI video generation. ModelFit's home page does the same matching for language models.

FAQ

Can a 16 GB Mac run Wan 2.2?

Only the FastMetal 5B derivative is listed for 16 GB, and no timing on a 16 GB Mac was found. Its peak was 9.34 GiB on a 36 GB M4 Max and 9.54 GiB on average on a 24 GB M5 Air. The full 5B model through mlx-video reached a 56.4 GB peak, and A14B needs both experts.

Does Wan 2.2 A14B run on a Mac at all?

Yes, through PocketAI's MLX packages. The Q4 pack took 10:25 end to end at 512×320, 81 frames and 10 steps on an M5 Max with 128 GB, with a 16.37 GiB peak. PocketAI lists 24 GB for short low-guidance tests. No 480P or 720P A14B Mac timing was found.

Which GGUF should I download for Wan 2.2 A14B?

Pick one quant and download its HighNoise and LowNoise files together. Q4_K_M is 19 GB for the pair and Q8_0 is 31 GB. One ComfyUI commenter reported Q5_K_M working on a Mac after the fp8 template failed, but no Mac timing or quality test was found for any GGUF quant.

What license does Wan 2.2 use?

Wan's README says the models in its repository are licensed under Apache 2.0, which allows commercial use under its terms. The QuantStack and PocketAI cards also list apache-2.0, and QuantStack adds that the original licensing terms and usage restrictions remain in effect. FastMetal is a separate derivative, so read its own card before using it.

Sources

• Wan2.2 README (architecture, 27B total and 14B active, license)
• Wan2.2 shared config (16 fps, 81 frames)
• Wan2.2 A14B T2V config (40 steps)
• Wan2.2 TI2V 5B config (24 fps, 121 frames, 50 steps)
• Wan-AI/Wan2.2-T2V-A14B model card (80GB VRAM note)
• Wan-AI/Wan2.2-I2V-A14B model card
• Wan-AI/Wan2.2-TI2V-5B model card (24GB VRAM note)
• Hugging Face file listing, official A14B T2V repo (exact bytes)
• Hugging Face file listing, official 5B repo (exact bytes)
• Hugging Face file listing, Comfy-Org/Wan-2.2 (exact bytes)
• QuantStack/Wan2.2-T2V-A14B-GGUF
• QuantStack T2V GGUF file listing (exact bytes)
• QuantStack I2V GGUF file listing (exact bytes)
• QuantStack 5B GGUF file listing (exact bytes)
• PocketAiHub/Wan2.2-T2V-A14B-MLX (benchmark table, requirements)
• PocketAI Q4 revision file listing (exact bytes, last modified)
• PocketAI Q8 revision file listing (exact bytes)
• FastMetal-QAD, Hao AI Lab, 2026-08-10 (M4 Max and M5 Air timings)
• FastVideo Apple Silicon install guide (16 GB tier)
• FastVideo/FastMetal-5B-QAD file listing (exact bytes)
• Generating AI Videos Locally: Wan 2.2 on the Mac, rotecodefraktion, 2026-10-10 (M3 Max timings)
• Blaizzy/mlx-video Wan README
• Rapid-MLX README (M4 Pro 48 GB benchmarks)
• ComfyUI Wan 2.2 5B workflow template
• ComfyUI Wan 2.2 14B T2V workflow template
• ComfyUI-GGUF README
• ComfyUI issue 9255, Float8 on MPS
• ComfyUI issue 15010, Wan 2.2 5B banding on Apple Silicon
• ComfyUI issue 14837, silent MPS attention corruption