DeepSeek-V4.1-Flash-1017 Maya GGUF

DeepSeek-V4.1-Flash-1017 Maya GGUF

Runs a 748B DeepSeek mixture-of-experts model on a home PC through two GGUF quantizations made for Project Maya. For people with a big-RAM machine and a fast SSD.

Made by peasantsmith 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 GGUF quantizations of DeepSeek-V4.1-Flash-1017, made for Project Maya. The model is a mixture of experts with 384 routed experts per layer and 6 active per token. Project Maya puts the experts that get called most onto the GPU, and loads the rest on demand from RAM and the SSD. That lets the model run in far less memory than its size suggests.

Who it suits: Hobbyists and developers who want to run a huge open model on their own hardware, and who accept a large download.

Industry | Who uses it | How they use it | Example
Software | Developers | Serve a large local model with a compatible API | Point an agent tool at the local address
Research | AI enthusiasts | Test a big mixture-of-experts model on one PC | Compare the two quantizations on their own prompts
Self-hosting | Privacy-minded users | Keep prompts on their own machine | Run chats without sending text to a cloud

If your PC has little RAM and no spare SSD space, you can skip it.

Where to use it:

• A desktop PC with a GPU, plenty of RAM and a fast SSD
• Windows through the START-MAYA.bat launcher, or Linux and macOS through the setup script
• Tools that talk to an OpenAI-style or Anthropic-style API on your own machine

Use case | Who | What they get back
Local chat | Individuals | A dashboard on your own machine to talk to the model
App development | Developers | An API with streaming, tool calls and pictures
Quantization study | Researchers | Two quants to compare against the original

How to set it up:

• A PC with enough RAM and disk, about 347.5 GB of downloads for Maya-S or 280.9 GB for Maya-XS
• A GPU, since the setup compiles the engine for it
• Git, to fetch the Project Maya repository the card links

1. Clone the Project Maya repository.
2. Run the setup script with the model name and the quant you want, maya-s or maya-xs.
3. Let it check the PC, download the files, build the engine and start it.

./setup.sh --setup --model DeepSeek-V4.1-Flash-1017 --ds4-quant maya-s     # Windows: START-MAYA.bat
./setup.sh --setup --model DeepSeek-V4.1-Flash-1017 --ds4-quant maya-xs    # v1.1.22 or later

You can also fetch only the files with the Hugging Face tool.

pip install -U huggingface_hub
hf download peasantsmith/DeepSeek-V4.1-Flash-1017-Maya-GGUF

How to use it: After setup, the engine starts a dashboard on port 8080 of your own machine. The same address also serves an API compatible with the OpenAI and Anthropic formats, with streaming, tool calls and pictures.

Tool | What it does
setup.sh --setup | Checks the PC, downloads verified files, builds the engine and starts it
--ds4-quant maya-s | Picks the quantization closest to the original
--ds4-quant maya-xs | Picks the smaller one that fits a 128 GB PC

Example: You run the setup line with maya-xs on a 128 GB PC. Once it finishes, you open the dashboard on port 8080 and chat with the model, which is held in memory with its Engram tables read from the SSD.

For newcomers and experts: If you are new, pick Maya-XS if your PC has 128 GB of memory, and let the setup script do the rest. Expect a download of hundreds of gigabytes.

For experts, the experts were rounded with GPTQ-style error feedback, using an importance matrix per expert and calibration text in DeepSeek's chat format. Attention, shared experts, embeddings and the output stay at Q8_0, and the files are ordinary llama.cpp GGUFs.

DeepSeek's own FP4 and FP8 release is the source
Experts are quantized with error feedback
Engram tables are kept in Q6_K on the SSD
Project Maya runs the model across GPU, RAM and SSD

Good to know: The card measured both quants on held-out text against the original model. The original weights through Maya's engine score 96.4 percent, which is the engine's own rounding.

Title: Same top token as the original model
Original weights in Maya's engine: 96.4 percent
Maya-S: 92.6 percent
Maya-XS: 87.5 percent

Title: The two quantizations from the card
Property | Maya-S | Maya-XS
Model in memory | 178.2 GB | 115.3 GB
Download, all files | 347.5 GB | 280.9 GB
Next-token loss vs the original | +3.2% | +8.7%

• License: MIT, the same as DeepSeek's original model.
• Cost to run: no API key or account is named. You supply a large PC and a lot of disk space.
• Data it reaches: the setup downloads checksummed files and fetches the vision tower from DeepSeek's checkpoint. The API listens on your own machine.
• Not for: small machines. The Engram tables alone take 161.3 GB of SSD space.

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 5.1K
• Weights: about 1 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 212,220
• Likes: 66
• License: mit
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Quantized versions: 3
• Tags: gguf, imatrix, mixture-of-experts, deepseek, project-maya, text-generation

Links

• Model: https://huggingface.co/peasantsmith/DeepSeek-V4.1-Flash-1017-Maya-GGUF
• Files: https://huggingface.co/peasantsmith/DeepSeek-V4.1-Flash-1017-Maya-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download peasantsmith/DeepSeek-V4.1-Flash-1017-Maya-GGUF

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