DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF

DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF

An expert pruned and quantized GGUF of DeepSeek-V4.1-Flash-1017 from ISTA-DASLab, aimed at code and agent work, with transformer weights of 72.1 GB. Suits developers with large memory machines.

Made by ISTA-DASLab 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 compressed GGUF version of DeepSeek-V4.1-Flash-1017 from ISTA-DASLab, aimed at code, agent tool use, long context and vision. Half of the 384 experts in each layer are removed, and the rest keep the quantization of the 2.00 bits per weight IQ2_XXS build. The transformer weights take 72.1 GB, which the card counts as 1.05 bits per parameter against the original.

Who it suits: Developers with large memory machines who want a code focused DeepSeek model in a smaller footprint, and researchers who study pruning together with quantization.

Industry | Who uses it | How they use it | Example
Software | Developers | Run a code and agent model locally with less memory | Ask it to fix an issue in a repository
Research | Compression researchers | Study expert pruning combined with quantization | Compare retained scores with the base
Infrastructure | Hosting teams | Size a deployment of a huge model | Plan around 72.1 GB of transformer weights

Skip it for general purpose chat, because the card says pruning costs ability outside code, tool use, long context and vision, and points to the unpruned releases for general use.

Where to use it:

• Programs that can load GGUF files
• Machines with enough memory for about 72 GB of transformer weights
• Coding and agent tools that you connect to a copy you run yourself

Use case | Who | What they get back
Code generation | Developers | Code written from a task description
Repository fixes | Agent users | Multi step changes for real code issues
Long context and image prompts | Teams | Answers that use long inputs or pictures

How to set it up:

• Python and the Hugging Face tools
• Disk space for 72.1 GB of transformer weights, plus the 110.6 GB of Engram lookup tables, which may stay on disk
• A program that loads GGUF files

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Load the GGUF files in your program.

pip install -U huggingface_hub
hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF

How to use it: Load the files and send code and agent prompts. The card measured every accuracy number at reasoning effort 50, so use the same setting when you compare.

Tool | What it does
hf download | Copies the repository files to your computer
Reasoning effort 50 | The setting used for all scores in the card

Example: Give an agent tool a repository and a bug report, and ask the model for a fix. It should return code changes. The card reports that it keeps 94.8 percent of the base score on SWE-bench Verified.

For newcomers and experts: If you are new to this, know that the files are huge and the card calls the release experimental. Try a smaller model first to learn the workflow.

If you know the field, each layer has 384 routed experts and 6 fire for each token. The pruning removes 192 of them per layer. The original has 763.2B parameters in all, of which 551.9B sit in the transformer. The 280.1B parameters that remain average 2.06 bits.

A code or agent prompt goes in
The router picks from the remaining experts
The 72.1 GB transformer runs
The engram tables are read per token

Title: Coder against the base model
SWE-bench Verified base: 81.20 points
SWE-bench Verified Coder: 77.00 points
LiveCodeBench v6 base: 89.71 points
LiveCodeBench v6 Coder: 88.57 points

Good to know:

• The card calls this an experimental release and asks for feedback on abilities missing from its calibration data.
• Hugging Face's own count shows 197B parameters, while the card describes a 763.2B base model.
• License: mit.
• Cost to run: none named, only your own hardware.
• Data it reaches: the prompts you give it. The source names no account or outside service.
• Not for: general purpose use. The card points to the unpruned releases for that.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: gguf
• Parameters: 197B
• Weights: about 98.3 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 342,280
• Likes: 225
• License: mit
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Quantized versions: 3
• Tags: gguf, text-generation, image-text-to-text, vision, quantization, pruning, expert-pruning, mixed-precision, ist-daslab, deepseek, code, conversational

Links

• Model: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF
• Files: https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-Coder-GGUF

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