MiniMax-H3-minirun

MiniMax-H3-minirun

Weights from the MiniMax H3 release and from community repacks of it, pinned by revision in the table below, repacked byte for byte into the container format Minirun reads. Minirun streams a model from an external SSD on Mac and iPhone through a memory budget you set, rather than loading it into memory, so the model does not have to fit in the machine. The reader is open source: Sources/StorageCore/Container/ in nanguoyu/minirun-app.

What this is

A byte-preserving repack with AdaLN pruning and precomputed modulation. Nothing here was quantized, requantized, retrained or distilled:

• the quantized matrices are the upstream affine 8-bit codes, scales and biases at group 64, copied verbatim and only regrouped into new files;
• the dense tensors are the upstream bf16 and fp32 bytes, copied verbatim;
• the per-block AdaLN projections and the time embedder are absent, and precomputed modulation caches for a 50-step base schedule and a 4-step Turbo schedule stand in their place. Removing whole tensors leaves every tensor that remains byte-identical.

Every byte of weight data here is a byte of one of the sources below, at the pinned revision, in a different order. The exceptions are container headers, alignment padding, the manifests, and the modulation caches, which were computed from the official bf16 AdaLN and time-embedder tensors.

MODIFICATIONS.md states the changes in full and NOTICE carries the attribution.

Contents

112 directories, 778 data files, 64.0 GB in total, plus one manifest per directory giving each file's shape, element width, offsets and digests.

Run it with Minirun

Get the app. On a Mac, download Minirun.dmg. On an iPhone, install TestFlight and open testflight.apple.com/join/QuYTybjV on the phone; the source is at nanguoyu/minirun-app if you would rather build it.

Point it at this repository. In Settings → Storage, use Add a folder… to register a folder on an external NVMe drive. In Settings → Models, open Find Models, select this repository and press Download -- or point Minirun at a copy you already have. Run Verify all files.

What to expect. Minirun downloads, verifies and stores this container today, and shows it as Container ready; a runtime for it is on the way. The two other containers published this way, Kimi K3 and DeepSeek V4 Flash, are chat models and run now.

Copy and verify speeds change with the drive, the cable and the budget; the enclosure and cable matter more than the SSD — USB4 against a 10 Gb/s hub is about 3× on the same bytes — and the iPhone needs a powered dock.

Requirements. An Apple-silicon Mac on macOS 15 or later; an iPhone 15 Pro or later on iOS 18 or later; an external NVMe drive with room for the model.

minirun.dev · Docs · This model · GitHub

Provenance

License

The model weights redistributed here are governed by the MiniMax H3 Community License Agreement, reproduced verbatim in LICENSE and copied unmodified from MiniMaxAI/MiniMax-H3 at revision 6818f6c32d12b210915e44ad56a4228c2608f160.

“MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.”

That is the notice Section III.4 of the Agreement requires to accompany every distribution; NOTICE carries it, together with the attribution for every source in the table above.

Two components carry a licence of their own in addition. The 4-step Turbo LoRA and the tiny approximate video decoder are released by their authors under the Apache License 2.0, and both are also derivatives of the MiniMax H3 Works, so their use remains subject to the Agreement as well. The two licences apply together, not in the alternative. The Agreement separately records that the H3 encoder is Qwen3-VL-32B under the Apache License 2.0.

The Agreement in LICENSE is the governing text. Nothing in this card is a substitute for reading it, and it carries obligations -- among them a territorial restriction, an acceptable-use policy, and a separate authorization requirement for large commercial users -- that this summary does not restate.