Minirun 0.5

Minirun 0.5

DeepSeek V4.1 Flash chats on Mac and iPhone. The 517 GB container answers on both devices. On a MacBook Pro (M1 Pro, 32 GB) over a USB4 enclosure it runs at about 5 seconds a token at the 14.9 GB Balanced budget, where it holds all forty blocks and the output head and a token reads only the 4.5 GB of experts it routes to. On an iPhone 16 Pro over a powered dock it runs at about 21 seconds a token at the 1.9 GB floor, reading 11.7 GB a token. The same question produced the same eighteen tokens at every budget and on both devices.

The memory dial reaches V4.1. Floor, Balanced and Generous are positions on the model's own ladder now: on a Mac a new chat starts at Balanced with everything held; on a phone it starts at Floor and the dial says how far Balanced is out of reach.

Downloads finish verified. What the transfer proved, file by file, is kept, so a finished download reads fully verified without a third pass over the drive. One copy is one row; a stopped transfer can be forgotten; a download no longer empties a repository into the folder you picked.

The model page, Settings, Storage, About, the models list and the download screens have been redesigned to read like product pages: hierarchy instead of cards, the copies on your drive above the facts about the model.

Answers render Markdown, the Instruments footprint row states one basis and names the process before and after a run, and every control has a name for VoiceOver. On iPhone, New chat opens the chat from anywhere and toolbar buttons stay inside the screen.

Minirun 0.4 (2026081801)

Minirun 0.4

DeepSeek V4 Flash is about a third faster on a Mac. The 64 sparse-attention heads are computed in one batch instead of one at a time. Summing them together rounds differently from summing them one by one, so an answer may be worded very slightly differently from 0.3 — and it is still byte-identical at every memory budget, which is the guarantee the memory dial rests on.

iPhone. Verify all files no longer ends the app on iOS 27, where the system rejected the background-continuation handler the app registered by wildcard; each continuation is now registered by name, and a refusal is shown instead of raised. The app targets iPhone only, which is what it claims, and this is the first release on TestFlight.

The streaming caret follows the text instead of sitting at the end of the first line.

Minirun 0.3 (2026081701)

Minirun 0.3

DeepSeek V4 Flash is about twice as fast on a Mac. Expert reads for all three projections start the moment the router names them, the expert gather is one dispatch per token instead of one per tile, projections are no longer forced one matrix at a time, the FP8 activation rounding uses half the operations, and graph-bounding evaluations no longer block. The answers are byte-identical to 0.2 at every budget.

Verified copies stay verified. When a model repository gains a commit that only touches its documentation, Minirun re-reads only the changed files and carries the rest of the evidence forward, and tells you how many files it read and how many it carried.

The memory-budget presets keep their names on one line.

Minirun 0.2 (2026081601)

The first Developer ID release of Minirun.

Download for Mac: https://downloads.minirun.dev/Minirun.dmg (signed, notarized; updates itself through Sparkle).
iPhone: build from this tag with Xcode — see the README.

• DeepSeek V4 Flash: memory dial with device-relative presets, resident layer cache and stated-scale execution; per-token phase attribution in Instruments; tile digests trusted under held verification authority.
• Kimi K3: sub-layer residency and layer-boundary reclaim, lowering the peak on iPhone.
• Verification continues in the background on iOS and finishes with a notification; interrupted verifications resume from a checkpoint.
• The memory-budget presets keep their names on one line.