DeepSeek-V4-Pro

DeepSeek-V4-Pro

Generates text with a 1.6T-parameter Mixture-of-Experts model that activates 49B per token and reads up to one million tokens. A DeepSeek preview for teams that can host about 1,599 GB of weights.

Made by deepseek-ai 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: DeepSeek-V4-Pro is a preview text generation model from DeepSeek. It uses a Mixture-of-Experts design with 1.6T parameters, of which 49B are active at a time. It reads up to one million tokens of context. A smaller sibling, DeepSeek-V4-Flash, has 284B parameters with 13B active.

Who it suits: Teams with serious hardware who want an open model for long inputs, coding and agent-style tasks. The weights take about 1,599 GB, so it is out of reach for an ordinary laptop or single graphics card.

Industry | Who uses it | How they use it | Example
AI research | Labs | Study long-context Mixture-of-Experts models | Test how the model behaves with an extra-long input
Infrastructure | Platform teams | Host a large open model on a cluster | Serve the weights to internal users
Software | Developers | Evaluate it on coding and agent tasks | Compare its output with another model on a coding job
Document work | Analysts | Feed in long material in one prompt | Ask questions about a large pile of text

You can skip it if you have a single machine or a small budget. The smaller Flash model in the same series is a lighter option.

Where to use it:

• Python with the transformers library, as listed by Hugging Face
• Large server setups on Windows, macOS or Linux that can hold the weights
• Your own infrastructure, since the card gives its own download links

Use case | Who | What they get back
Long-context reading | Analysts | Answers drawn from up to one million tokens
Coding | Developers | Code written by an open model
Agent work | Agent builders | A model for multi-step tasks

How to set it up: This needs a huge amount of storage and memory, so check the numbers first.

• Python with the Hugging Face tool
• About 1,599 GB of disk for the I8 weights
• Server hardware that can hold and serve a 1.6T-parameter model

1. Install Python if you do not have it.
2. Install the Hugging Face hub tool.
3. Check that you have room for the weights.
4. Download the repository.
5. Follow the card for how to serve it, which continues past what is shown here.

pip install -U huggingface_hub
hf download deepseek-ai/DeepSeek-V4-Pro

How to use it: The run steps are cut off in the part shown, so follow the full card once the weights are in place. The card names a maximum reasoning effort mode called DeepSeek-V4-Pro-Max.

Example: You give the model a long document and ask questions about it. The card says the series reads up to one million tokens, so the whole text can go into one prompt.

For newcomers and experts: A Mixture-of-Experts model has many specialist parts and uses only some of them for each token. That is why 1.6T parameters can run with 49B active. This one is meant for large servers, so a first small step is to read the card and the size figures before downloading anything.

The card describes a hybrid attention design that mixes Compressed Sparse Attention and Heavily Compressed Attention. At one million tokens it needs 27 percent of the single-token compute and 10 percent of the KV cache of DeepSeek-V3.2. It also uses Manifold-Constrained Hyper-Connections and the Muon optimizer. Pre-training used more than 32T tokens. Post-training trains domain experts with SFT and RL, then merges them by on-policy distillation.

Pre-train on more than 32T tokens
Train domain experts with SFT and RL
Merge them by on-policy distillation
Serve the weights for long prompts

Title: DeepSeek-V4-Pro at one million tokens, as a share of DeepSeek-V3.2
Single-token inference compute: 27%
KV cache: 10%

Good to know:

• The card calls this a preview version of the V4 series.
• Its claims about being the best open model and reaching top coding scores are the authors' own.
• The card says V4-Flash reaches reasoning like Pro with a larger thinking budget but trails on pure knowledge and the hardest agent work.
• License: mit.
• Cost to run: none named, but the hardware needs are huge.
• Data it reaches: files downloaded from Hugging Face, plus the prompts and documents you give it.
• Not for: small machines or casual local use, given the 1,599 GB of weights.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 1.6T
• Weights: about 1,599 GB in I8
• Formats: Safetensors
• Downloads in the last 30 days: 322,162
• Likes: 5,702
• License: mit
• Quantized versions: 31
• Fine-tunes: 17
• Tags: transformers, safetensors, deepseek_v4, text-generation, 8-bit, fp8

Links

• Model: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
• Files: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download deepseek-ai/DeepSeek-V4-Pro

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