← All models

DeepSeek V4.1 Flash

Multimodal chat and agents

Chat readyMaciPhone
Download for Mac Get it for iPhone

Mac: macOS 15 or later on Apple silicon, and it updates itself. iPhone: through TestFlight, iPhone 15 Pro or later on iOS 18.

  1. Download Minirun
  2. Models > Find Models — pick DeepSeek V4.1 Flash
  3. Verify all files
  4. New chat
The Minirun instruments panel on macOS at the end of a DeepSeek V4.1 Flash turn: 4.8 s per token, 4.51 GB read for that token, and a peak of 14.7 GB against the 14.9 GB budget.The same model on an iPhone 16 Pro: 0:21 / token, 11.7 GB read for that token, a peak of 971 MB against a 1.90 GB budget, and where the time went.
DeepSeek V4.1 Flash at the end of a turn, with the instruments Minirun shows while it answers.

About

DeepSeek V4.1 Flash — a 552B multimodal mixture-of-experts model with a million-token context. It holds a conversation on a Mac and on an iPhone 16 Pro, streaming the 517 GB container straight off your drive.

On a Mac it answers at about 5 s / token at the 14.9 GB Balanced budget, where it keeps all forty of its layers and its output head in memory and reads 4.5 GB of experts a token. It runs at 3.4 GB, streaming everything. On an iPhone 16 Pro: about 21 s / token at 1.9 GB, reading 11.7 GB a token. The first reply takes about half a minute to start on the Mac and about a minute on the phone. The pace depends on the drive, the cable and the budget.

What you need

About 3.4 GB of memory on Mac, 1.9 GB on iPhone 16 Pro, and 517 GB free on your SSD.

Model requirements
Technical details
Stored precision
FP4 experts · FP8 matrices and memory tables · BF16/F32 remainder
Container size
517 GB
License
MIT
Repository
nanguoyu/DeepSeek-V4.1-Flash-minirun
Source model
deepseek-ai/DeepSeek-V4.1-Flash
View on Hugging Face Read the docs
DeepSeek V4.1 Flash — Minirun