A 115GB download is the price of entry for running MiniMax-H3 on Apple hardware. The multimodal system, which accepts text, images, audio, and video and produces clips of up to 15 seconds with audio, now has an MLX port, and it has been tested on an M5 Max MacBook Pro.
Simon Willison tested the port on an M5 Max MacBook Pro. The model download is roughly 115GB. His video generation run took under 45 minutes, which puts the model in the realm of batch local work rather than interactive use.
What the MLX Port Actually Is
The package wraps MiniMax-H3 for MLX, Apple’s machine learning framework for Apple Silicon. That means the model runs on the unified memory of the M-series chips without a separate GPU.
MLX is Apple’s array framework for machine learning, and it matters here because it works directly with the unified memory architecture of M-series chips. Instead of copying weights into a separate VRAM pool, the model reads from the same memory the CPU and GPU share, which is why a 115GB model can run at all on machines with enough unified memory. The port ships with several quantization options, and the practical rule is to pick the highest precision your free memory allows, because video generation is where quantization artifacts become visible in faces and fine text. On the M5 Max that Simon tested, short clips complete in minutes rather than tens of minutes, though the Metal GPU falls back to slower paths under memory pressure, so background apps cost you generation time.
The practical significance is straightforward. MiniMax-H3 was announced as a flagship multimodal system, and most local-run claims for models like this end up being theoretical. Here there is a working port with a documented test on real hardware.
The numbers matter for planning. 115GB of model files is a serious download and a serious disk footprint. The M5 Max with unified memory handles it, but this is not a casual install, and generation is measured in tens of minutes, not seconds.
How It Compares to the Native Video Race
MiniMax H3 ships with native dual-channel audio, which is a step beyond video models that produce silent clips and need a separate audio pass. Running that pipeline locally on Apple hardware is genuinely new.
In practice, video quality sits in the same family as the cloud-hosted H3, with the trade-offs you expect from a local run. Faces stay stable in short clips, motion is smoother than the previous generation of open video models, and the native audio track means dialogue and effects come out in one pass instead of a separate audio pipeline. The honest limits: very long clips push memory pressure hard, scenes with many moving objects occasionally produce artifacts, and the first generation after loading pays a one-time warm-up cost. None of this is unusual for a 2026-era local multimodal model, but it is worth knowing before you commit a weekend to a 115GB download.
For context, the broader AI video market in 2026 is dominated by cloud APIs. ByteDance’s Seedance 2.5, PixVerse, and others push strong models through hosted endpoints. A local option matters for confidentiality-sensitive work, offline workflows, and anyone who wants to avoid per-generation API costs.
The company behind H3 is pushing the same open-weights approach into music. MiniMax Music 3.0, released August 14, pairs an 8B language model with a 2.4B audio renderer and generates a complete song in one pass, with open weights following the same playbook as H3.
The Cost and Performance Reality
Here is the honest part. Running a 115GB multimodal model locally is not a convenience play. The download, the disk space, and the generation time all point to serious hardware and patience.
The value is in what it enables: full control over generation, no per-clip API fees, and data that never leaves the machine. For studios, researchers, and teams with sensitive content, that tradeoff is often worth it.
The same logic extends beyond video. NVIDIA’s Nemotron 3.5 Lightning, an open mixture-of-experts model built for local agents, activates only 3B of its 30B parameters, another sign that the local-run market is moving toward smaller active footprints. If you are evaluating local-agent models like Nemotron 3.5 Lightning, it complements this multimodal setup rather than competing with it.
The local video picture is changing fast at the other end of the size spectrum. FastMetal-QAD, released the same week, runs a 5B video model in under a minute on a 16GB Mac, a fraction of the memory H3 demands, while Qwen 3.8 27B covers vision and reasoning in a 17GB file. The trade-off is capability: H3 remains the heavier option for full multimodal generation with native audio, and the smaller models are the practical choice for quick clips on everyday laptops. Choosing between them is a memory-versus-quality decision, and our FastMetal on Mac and Qwen 3.8 27B pieces cover both ends of that spectrum in detail.
Getting Started
If you want to try it, the path is the standard MLX workflow: install the package, download the model weights, and run generation with the documented commands. Expect the download to take a while, and expect the first generation to be slow while the model loads.
Who should actually download it: developers building local video pipelines who need full control over generation parameters, teams with privacy requirements that rule out cloud APIs, and researchers comparing open multimodal models. Who should skip it: anyone with less than 128GB of unified memory, anyone who needs long high-resolution renders, and anyone whose hardware budget would be better spent on the smaller local models described above.
The practical recommendation: test with a short clip first, verify the output pipeline end to end, and only then commit to longer generation runs.
The bottom line: MiniMax-H3 on Apple Silicon is a flagship-local experience with a flagship-local price. It is the right tool when you need the full multimodal package, native audio, and total control, and you have the memory to host it. For everyone else, the fast-growing family of small local video models is catching up quickly, and the decision increasingly comes down to how much quality you need versus how much RAM you have.