DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw

DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw

Runs a 261B DeepSeek V4.1 Flash model packed to about 1.75 bits per weight for the sushi engine on Apple Silicon Macs, with image input and up to 1,048,576 tokens of context.

Made by beamster on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds DeepSeek-V4.1-Flash-1017 repacked by beamster for sushi, an inference engine built for Apple Silicon. Revision 2.0 stays closer to the original model than revision 1.0 in every test category. It also stores the expert weights in an order that sushi can stream from the SSD into its expert cache.

Who it suits: It suits Mac owners with Apple Silicon and plenty of disk space who want to run a huge model locally through sushi.

Industry | Who uses it | How they use it | Example
Local AI hobby | Mac enthusiasts | Run a huge open model on Apple Silicon | Trying a 261B model on a high-end Mac
Software development | Developers | Test tool calling and agent behavior locally | Running a tool-use benchmark against the model
AI research | Quantization researchers | Compare compressed packs against the original checkpoint | Reading the KLD tables for each category
Long documents | Analysts | Feed in long inputs, up to the stated context limit | Asking questions about a long report

You can skip it if you use Windows, Linux or an Intel Mac, since this pack works only with sushi on Apple Silicon.

Where to use it:

• The sushi engine, version 1.5.0 or later, on Apple Silicon
• macOS 26.2 or later
• Any computer with Python, only to download the files

Use case | Who | What they get back
Local chat and image questions | Mac users | Text answers, with a vision tower for image input
Tool calling | Developers | Tool-use behavior, measured at 92 out of 100 on revision 1.0
Long-context work | Analysts | A context window of up to 1,048,576 tokens

How to set it up: You need a recent Apple Silicon Mac and about 309 GiB of disk for the pack.

• An Apple Silicon Mac on macOS 26.2 or later
• The sushi engine, version 1.5.0 or later, since older versions refuse this pack by name
• About 309 GiB of disk, made of 120 GiB of weights and 189 GiB of Engram tables
• Python and huggingface_hub if you download with the Hugging Face tool

1. Install sushi 1.5.0 or later on your Mac.
2. Download the pack with the Hugging Face tool, or with sushi pull if you already hold revision 1.0.
3. Follow the sushi usage guide for launch flags, memory and context on your Mac.

pip install -U huggingface_hub
hf download beamster/DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw

How to use it: You run the model through sushi, whose usage guide holds the launch flags. If you stay on an older sushi, you can fetch the first revision by its tag instead.

hf download beamster/DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw --revision v1.0

Tool | What it does
sushi pull | Fetches revision 2.0 for someone who holds 1.0, changing only the files that differ
hf download with --revision v1.0 | Fetches revision 1.0 for an older sushi

Example: You already hold revision 1.0 and run sushi pull. It compares file contents, downloads only the files that changed, and checks each one before it replaces the old copy.

For newcomers and experts: If you are new to this, it is a giant model squeezed so it fits a Mac, and sushi is the program that runs it. A first small try is to read the sushi usage guide and check what memory your Mac needs.

For people who know the field, routed experts use STQ, a trellis quantization with paired codewords, at an average of 1.75 bits. The trunk uses 4 bits and the output head 6 bits. The Engram tables stay as released and are read from the SSD, never loaded into GPU memory. The pack also includes the three DSpark speculative-decoding stages.

Title: KLD against the original model, lower is better
Revision 2.0: 0.1576 KLD
Revision 1.0: 0.1843 KLD
oQ2.5e: 0.1717 KLD

You download the pack
sushi loads the trunk and 1.75-bit experts
Engram tables stream from the SSD
You get text answers, or image answers via the vision tower

Good to know: The quality numbers come from the pack's author and cover eight categories with two prompts each, so treat them as a guide rather than a benchmark.

• License: MIT, as a derivative of the DeepSeek-AI base model
• Cost to run: none named
• Data it reaches: files on your disk, with Engram tables read from the SSD
• Not for: sushi versions before 1.5.0, or machines other than Apple Silicon Macs

On Hugging Face, 13 November 2026

• Task: image text to text
• Parameters: 261B
• Weights: about 261 GB in F8_E4M3; about 131 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,478
• Likes: 23
• License: mit
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Tags: safetensors, deepseek_v41, sushi, apple-silicon, image-text-to-text, fp8

Links

• Model: https://huggingface.co/beamster/DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw
• Files: https://huggingface.co/beamster/DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download beamster/DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw

Then follow the usage in its model card. Nothing in this kit runs by itself.