Kimi-K3-minirun

The weights of moonshotai/Kimi-K3 at revision 9f62e4e9fffbd0a83ddd60e1c209d828994b3569, repacked byte for byte into the container format Minirun reads. Minirun streams this 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. No requantization, no retraining, no numerical change of any kind:

• the routed expert weights are the upstream MXFP4 packed values and their e8m0 scales, copied verbatim and only reordered;
• all non-expert tensors are the upstream BF16/F32 bytes, copied verbatim.

The container files are therefore a rearrangement of the source bytes. Every value in this repository originates from the pinned revision above.

Contents: 93 layer directories, 372 container files, 1.56 TB total, plus a per-layer manifest describing offsets and shapes. Vision tensors (vision_tower.*, mm_projector.*) are not included; this is the text stack only.

Run it with Minirun

Get the app. On a Mac, download Minirun.dmg. On an iPhone, build it from nanguoyu/minirun-app; the README there has the steps.

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, then start a chat.

What to expect.

On disk	1.56 TB
MacBook Pro (M1 Pro, 32 GB, USB4 enclosure)	≈ 70 s / token at an 8 GB budget
iPhone 16 Pro (USB 3 port, powered dock)	≈ 220 s / token at a 5.8 GB budget, replies capped at 2 tokens per turn

The iPhone figure is a demonstration that the size ceiling is gone, not a way to chat. 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

Source model	moonshotai/Kimi-K3
Source revision	9f62e4e9fffbd0a83ddd60e1c209d828994b3569
Relationship	byte-preserving repack (no requantization)

License

This repository redistributes model weights owned by Moonshot AI under the Kimi K3 License, reproduced verbatim in LICENSE and copied unmodified from the source repository at the pinned revision above.

Copyright (c) 2026 Moonshot AI.

Two clauses of that licence apply to anyone using these files downstream, and are highlighted here rather than left to be discovered:

• §2 — Model as a Service. If you (or your affiliates) operate a Model as a Service business and your aggregate revenue exceeds US$20 million over any consecutive 12 months, you must enter into a separate agreement with Moonshot AI before any commercial use of these weights or works derived from them.
• §3 — Attribution on the interface. If you use these weights in a commercial product or service with more than 100 million monthly active users or more than US$20 million in monthly revenue, "Kimi K3" must be prominently displayed in that product's user interface.

§4 exempts purely internal use. The weights are provided "as is", without warranty of any kind, per §5. Refer to LICENSE for the governing text; the summary above is not a substitute for it.