DeepSeek-V4.1-Flash-1017 iQ-MLX Pruned192 2.25bpw

DeepSeek-V4.1-Flash-1017 iQ-MLX Pruned192 2.25bpw

Runs a pruned, 2-bit DeepSeek-V4.1-Flash on a 128 GB Mac through mlx-serve, keeping 192 of 384 experts and a 1M-token context. Strongest on code, reasoning and tool calls.

Made by ddalcu 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 is a pack of DeepSeek-V4.1-Flash-1017 built for mlx-serve on a Mac with 128 GB of memory. In each layer, half of the 384 routed experts stay, 192 in all, and GPTQ compresses them to 2 bits. The attention trunk and the shared expert are left exactly as in the original release. The pack keeps the 1M-token context, speculative decoding and the vision tower.

Who it suits: It suits people with a 128 GB Mac who want a large coding and agent model running locally, and researchers studying expert pruning.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Software development | Developers | Run a large coding and reasoning model on their own Mac | Asking it to work through a coding task locally
Local agents | Agent builders | Use it for tool calls and agent sessions | An agent that calls tools without a cloud service
Model research | Researchers | Study what pruning half the experts costs | Reading the KLD and top-1 tables by slice

People who need non-English text, chat quality, or a Mac with less than 128 GB can skip it. With 256 GB, the card says to use the 3.0 bpw pack.

Where to use it:

• mlx-serve, version 26.11.2 or later, on a Mac
• The Mac's internal SSD, which the card says should hold the model
• Windows, macOS or Linux: you can download the files with Python, though the pack is made for a Mac

Title: Jobs it does
Use case | Who | What they get back
Coding help | Developers | Code written by a model that holds up well after pruning
Tool use and agents | Agent builders | Tool calls and agent sessions that the card tested
Reasoning tasks | Researchers | Reasoning answers that stay close to the released model

How to set it up: You need these first:

• A Mac with 128 GB of memory, since the pack uses about 92 GB resident
• About 296 GB of free space on the internal SSD for the download
• mlx-serve 26.11.2 or later, which the card links through the MLXServe app

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Install mlx-serve 26.11.2 or later.
4. Start the server with the card's command.

pip install -U huggingface_hub
hf download ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-Pruned192-2.25bpw

The download includes 203 GB of Engram tables. They are read from the SSD for each token and do not stay in memory.

Title: Which pack for which Mac
Pack | Memory in use | Download | Fits
Lossless | 288 GB | 477 GB | A 512 GB Mac, or as a reference
iQ-MLX-3.0bpw | 209 GB | 427 GB | A 256 GB Mac
iQ-MLX-Pruned192-2.25bpw, this pack | 92 GB | 296 GB | A 128 GB Mac

How to use it: Start the server with the card's command, which names this model and turns serving on.

mlx-serve --model ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-Pruned192-2.25bpw --serve

Each release is a tag on the repository, and main is the newest. To pin one, add the revision flag to the download command.

hf download ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-Pruned192-2.25bpw --revision v1.0

Example: You start the server on a 128 GB Mac and send it a coding task. According to the card, code, reasoning and tool calls survive the pruning well, and chat and non-English text suffer more.

For newcomers and experts: If you are new to this, know that this is a huge download meant for one kind of Mac. Check that you have 128 GB of memory and enough SSD space before you start.

If you work in the field, note that 6 experts are active per token, as in the release. The kept experts were chosen by a blend of routing-weighted output energy and the release's own routing over real agent sessions. The kept experts carry 82% of the router's weight on agent traffic. A previous pick based on output energy alone had a better KLD score yet got stuck in loops during long agent sessions.

Download the pack and the Engram tables
Install mlx-serve 26.11.2 or later
Start the server with the card's command
Tool, code and reasoning requests run locally

Good to know: Pruning is not free. The card scores this pack against the released model's own outputs, and top-1 agreement is 85.82% overall, against 92.64% for the 3.0 bpw pack.

Title: Top-1 match with the released model, by slice
Tools: 92.0%
Reasoning: 91.9%
Code: 88.2%
Agent: 87.9%
Chat: 79.6%
Multilingual: 65.4%

• License: mit
• Cost to run: none named, apart from needing a Mac with 128 GB of memory
• Data it reaches: the text you send it, with the Engram tables read from your SSD
• Not for: non-English text, which degrades badly, or Macs with less than 128 GB of memory

On Hugging Face, 13 November 2026

• Task: text generation
• Library: mlx-serve
• Parameters: 27.1B
• Weights: about 109 GB in U32
• Formats: Safetensors
• Downloads in the last 30 days: 606
• Likes: 4
• License: mit
• Languages: en, zh
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Tags: mlx-serve, safetensors, deepseek_v41, gptq, expert-pruning, engram, dspark, text-generation

Links

• Model: https://huggingface.co/ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-Pruned192-2.25bpw
• Files: https://huggingface.co/ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-Pruned192-2.25bpw/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ddalcu/DeepSeek-V4.1-Flash-1017-MLX-Serve-iQ-MLX-Pruned192-2.25bpw

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