DeepSeek-V4.1-Flash-minirun

The weights of deepseek-ai/DeepSeek-V4.1-Flash at revision df42c109f1defefcbfcedbe7d905718a12266e40, 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 FP4 packed values and their ue8m0 scales, copied verbatim and only reordered;
• the remaining quantized matrices are the upstream FP8 e4m3 values and their ue8m0 32x32 block scales, copied verbatim;
• the two conditional-memory tables are the upstream FP8 e4m3 rows and their own ue8m0 scales, copied verbatim and regrouped so that a row and its scales sit together;
• all other tensors are the upstream BF16/F32 bytes, copied verbatim.

Every byte of weight data in this repository is a byte of the source checkpoint at the pinned revision above, in a different order. The only bytes that are not are container headers and zero padding: where a container's scale region is shorter than its alignment unit, and at the tail of each row page.

Contents: 44 directories, 557 data files, 517.3 GB total, plus a per-directory manifest giving each file's shape, element width and offsets.

Layout

Three kinds of data file, all described completely by the per-directory manifest.json beside them.

Suffix	Holds	Addressed by
.mxfp4tile	routed experts, two per tile	tile index; expert_order maps an expert id to its tile and half
.fp8tile	one FP8 matrix and its 32x32 scale grid	one tile
.engrampage	a slice of a conditional-memory table	row number

An .engrampage file is the one worth describing here, because it is not a matrix. The 2 conditional-memory tables hold 768,022,850 rows between them, stored as fixed 4096-byte pages of 15 rows each: every row's 256 value bytes, then every row's 8 scale bytes, then zero padding. A row and its own scales are therefore always inside one aligned 4096-byte read, which the two separate upstream runs could not offer. The manifest gives the page size, the rows per page, the offset of the scale region inside a page, and which slice of the table each file covers, so a row's address is arithmetic:

part  = the file whose row_base <= row < row_base + rows
page  = (row - row_base) // rows_per_page
at    = first_page_offset + page * page_bytes
value = at + (row - row_base) % rows_per_page * row_weight_bytes
scale = at + scale_offset_in_page + (row - row_base) % rows_per_page * row_scale_bytes

Source files

The model configuration, the tokenizer, the tokenizer-normalisation code and its test vectors, the upstream card and the technical report are verbatim files from the same pinned source revision. Every one of their byte counts, SHA-256 identities and original upstream paths is recorded in index.json under source_files.

They are all repository root files, including inference-config.json — the argument file DeepSeek's own reference code loads, upstream inference/config.json. A directory would have made them look like undeclared model payload to anything that counts this repository's files against what index.json declares, so the upstream directory became a name prefix and the original path is in the table instead.

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 Minirun does with it today. It downloads this repository, verifies every file against the published tree, keeps it on the drive you chose, and chats with it from there. On a MacBook Pro (M1 Pro, 32 GB) over a USB4 enclosure it answers at about 5 seconds a token at the 14.9 GB Balanced budget, holding all forty blocks and the output head; on an iPhone 16 Pro over a powered dock, about 21 seconds a token at the 1.9 GB floor. The same question gives the same eighteen tokens at every budget and on both devices.

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 517 GB.

minirun.dev · Docs · This model · GitHub

Provenance

Source model	deepseek-ai/DeepSeek-V4.1-Flash
Source revision	df42c109f1defefcbfcedbe7d905718a12266e40
Relationship	byte-preserving repack (no requantization)

License

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

Copyright (c) 2023 DeepSeek.

The MIT License permits use, copying, modification and redistribution, including commercially, provided the copyright notice and the permission notice are included in all copies or substantial portions of the Software. The weights are provided "as is", without warranty of any kind. Refer to LICENSE for the governing text; the summary above is not a substitute for it.