DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp

DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp

A mixed-precision MLX copy of DeepSeek-V4.1-Flash-1017 that keeps the Engram tables at FP8. Made for people with hundreds of GB of memory who want to run the model with MLX tools.

Made by mlx-community 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 variant of the oQ4e-mtp MLX checkpoint of DeepSeek-V4.1-Flash-1017. Here the n-gram Engram tables stay at their original FP8 precision, while all other weights match the other copy. The original model is a multimodal mixture of experts whose backbone holds 552B parameters. The full checkpoint averages 5.366 bits per weight.

Who it suits: It suits advanced users and teams with huge amounts of memory who want to run a big multimodal model through MLX and keep the Engram tables at higher precision.

Industry | Who uses it | How they use it | Example
AI infrastructure | Platform engineers | Host a huge open model on MLX hardware | Serving the model to internal tools
Research | ML researchers | Compare Engram tables kept at FP8 with 4-bit ones | Checking differences between the two copies
Software | Advanced developers | Test coding and agentic behavior locally | Trying a coding task with the model
Multimodal work | Applied teams | Combine text and image inputs in one model | Asking a question about a picture

Skip it if you have a normal laptop or desktop, since the stored tensors take about 477 GiB.

Where to use it:

• MLX based tools
• Machines with hundreds of GB of memory, or with fast disk for the Engram offload
• Windows, macOS and Linux, with Python, to download the files

Use case | Who | What they get back
Reasoning and coding | Developers | Text replies from the full model
Image and text input | Applied teams | Answers that use pictures
Offload experiments | Researchers | Same resident memory with Engram on SSD

How to set it up:

• Python, to run the download command
• About 476.751 GiB of storage for the tensors, according to the card
• MLX tooling that can load the checkpoint

1. Install the Hugging Face tools.
2. Download the model files to a drive with plenty of space.
3. Load the checkpoint with your MLX tooling.

pip install -U huggingface_hub
hf download mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp

How to use it: The part of the card available here gives no run commands, only how the model was quantized. Load the checkpoint in MLX tooling that supports it. The card says memory stays the same as the sibling copy when the Engram tables are offloaded to SSD.

Component | Storage
Language backbone | 300.132 GB
Engram tables | 202.758 GB
Vision encoder | 0.971 GB
MTP head | 8.047 GB

Example: You load the checkpoint in an MLX tool on a machine with enough memory and ask a coding question. The model answers with text. The card shows no sample run, so treat this as the general idea.

For newcomers and experts: This is a giant model in a large file, meant for people with huge amounts of memory. If you are new, start with a small model and come back to this one later.

If you know the field, the Engram tables are stored as MXFP8 in groups of 32, so each group shares one E8M0 scale over E4M3 values. The other weights are as in the oQ4e-mtp copy. Without the Engram tables the average is 4.365 bits per weight.

Start from the oQ4e-mtp copy
Keep Engram tables at original FP8
Leave other weights unchanged
Save the MLX checkpoint

Title: Storage by component, in GB
Language backbone: 300.132 GB
Engram tables: 202.758 GB
Vision encoder: 0.971 GB
MTP head: 8.047 GB

Good to know:

• Size: 763.205B logical weights in total, with about 476.751 GiB of tensor storage.
• Difference: the Engram tables use 8.25 bits per weight here, against 5.00 in the 4-bit sibling.
• Source: a quantized copy of DeepSeek-V4.1-Flash-1017.
• License: MIT, following the source model.
• Cost to run: none named, though the memory needs are huge.
• Data it reaches: it downloads model files from Hugging Face and works on the text and images you give it.
• Not for: ordinary laptops or small desktops.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: mlx
• Parameters: 289B
• Weights: about 289 GB in F8_E4M3; about 144 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 3,238
• Likes: 13
• License: mit
• Based on: deepseek-ai/DeepSeek-V4.1-Flash-1017
• Tags: safetensors, deepseek_v41, quantized, image-text-to-text

Links

• Model: https://huggingface.co/mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp
• Files: https://huggingface.co/mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-ngram-mtp

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