Best video models for a 24 GB Mac
Which open-weight video models run on a 24 GB Mac. Verdicts come from exact Hugging Face file sizes, model card and app maker statements, and published runs.
- Run
- 4
- Unknown
- 1
- Does not fit
- 1
4 of 6 listed video models run on a 24 GB Mac: MiniMax H3, LTX-2.3, Wan2.1 and Wan2.2 TI2V 5B. Highest blind-vote rank among them: MiniMax H3, #8 of 48 in text to video (Arena, snapshot 2026-09-22). No figure covers Mochi 1 preview at this size yet. Wan2.2 A14B needs more memory.
License check before commercial use: MiniMax H3, LTX-2.3.
Get the weekly local AI newsletter
The Thursday ModelFit newsletter: new local models and what fits your machine.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Which video models run on a 24 GB Mac?
6 modelsRuns means a published run on 24 GB or less, or a model card or app maker figure at or below it. Does not fit means the weights the app keeps resident in the smallest complete build are larger than 24 GB. A Mac has one memory pool, so the text encoder cannot move out to a separate system RAM; an app that loads the text encoder and the main model one after the other counts only its largest stage. unknown means no source covers this size yet; it is not a failure.
| Model | Verdict | Why | Quality | App |
|---|---|---|---|---|
| MiniMax H3MiniMax H3 Community License Agreement | Runs · Published | A published run on M4 Pro · 24 GB used the Argus calibrated INT8 DiT with the official text encoder and VAEs build (115.5 GB download) in Argus-AiTeam/minimax-h3-mac (MLX streaming).Argus MLX route README (INT8 run) | #8 of 48 in text to video11,417 votes | Argus-AiTeam/minimax-h3-mac (MLX streaming), antirez/h3.c (Metal), ddalcu/mlx-serve, stable-diffusion.cpp, ComfyUI |
| LTX-2.3LTX-2 Community License Agreement | Runs · Published | A dated run on Apple M5 Pro with 24 GB of unified memory: ltx-2-mlx-swift int8 streamed, 704x512, 121 frames.ltx-2-mlx-swift README | No independent score yet | Lightricks/LTX-2 ltx-pipelines, appautomaton/ltx-video-mlx, ComfyUI |
| Wan2.1Apache License 2.0 | Runs · Stated | Stated by the source: 16 GB of unified memory runs it, below this 24 GB tier (mlx-video 4-bit conversion).mlx-video Wan README | No independent score yet | ml-explore/mlx-examples Wan2.1, ComfyUI, FastVideo MLX (FastMetal) |
| Wan2.2 TI2V 5BApache License 2.0 | Runs · Published | A published run on M5 · 24 GB used the FastMetal 5B QAD (DMD2-distilled INT8 derivative, not the base model) build (19.5 GB download) in FastVideo MLX (FastMetal).FastMetal-QAD announcement | No independent score yet | Blaizzy/mlx-video, FastVideo MLX (FastMetal), ComfyUI |
| Mochi 1 previewApache License 2.0 | unknown | The smallest complete build, calcuis GGUF Q3_K_M with Q4_0 T5 and FP8 VAE, is a 7.7 GB download, but no stated figure or published run covers 24 GB of unified memory yet. | #48 of 48 in text to video5,893 votes | Diffusers, ComfyUI |
| Wan2.2 A14BApache License 2.0 | Does not fit est. | The smallest complete build, QuantStack GGUF Q4_K_M (high and low noise), keeps 26.3 GB of weights resident, more than 24 GB of unified memory. | #42 of 48 in text to video10,458 votes | PocketAI Wan2.2 MLX (pocketai-wan-video), ComfyUI |
Quality: blind human votes from Arena (arena.ai), CC BY 4.0, snapshot 2026-09-22. Scores are for the Arena checkpoint, not your local quantized build.
Frequently asked questions
What is the best video model for a 24 GB Mac?
MiniMax H3 has the highest blind-vote rank among the video models that run on a 24 GB Mac: #8 of 48 in text to video, from 11,417 votes (Arena, snapshot 2026-09-22). LTX-2.3, Wan2.1 and Wan2.2 TI2V 5B also run.
Which video models run on a 24 GB Mac?
MiniMax H3 (published run on M4 Pro · 24 GB), LTX-2.3 (run on Apple M5 Pro with 24 GB), Wan2.1 (16 GB stated by mlx-video Wan README) and Wan2.2 TI2V 5B (published run on M5 · 24 GB).
Which video models need more than 24 GB on a Mac?
Wan2.2 A14B (smallest complete build QuantStack GGUF Q4_K_M (high and low noise), 26.3 GB). Its resident weights alone are larger than the memory.