DeepSeek Open Models: Model List, Licenses and Articles

September 26, 2026September 27, 2026

About DeepSeek

DeepSeek (深度求索) is a Chinese company founded in 2023. Its deepseek-ai organization on Hugging Face describes it as a company “dedicated to making AGI a reality".

DeepSeek’s open models are known for releasing very large MoE (Mixture-of-Experts) models, weights included, under the MIT license. Total parameter counts run from several hundred billion to 1.6T, but only a fraction (billions to tens of billions) is active per token, which keeps compute down.

What It Releases

The list of recent models (release date, task, license, parameters) is in the “Recent Models on Hugging Face" table below, collected by code every day. Here are the main points.

• Long-context efficiency is at the center of the design. Both the V4 preview models and V4.1-Flash support a one-million-token context. The V4 model card says a hybrid of two compressed attention mechanisms (CSA and HCA) brings per-token inference compute down to 27% and the KV cache to 10% of DeepSeek-V3.2 at a one-million-token context.
• Weights ship in low precision from the start. V4-Pro and V4-Flash are released in mixed precision, with MoE expert parameters in FP4 and most other parameters in FP8 (the Base versions are FP8).
• V4.1-Flash moves into multimodal models. DeepSeek first released V4-Flash-Vision-Exp as “our first experimental multimodal model in the DeepSeek-V4 family"; V4.1-Flash is trained on images and text together from the start. Its model card gives the backbone as 552B and lists components such as the Engram memory (196B) separately; the parameter count in our article (763.2B) is the Hugging Face total including them.
• For benchmark tables and architecture details, follow the family pages from the “By Model Family" table below.

Licenses

Every V4 and V4.1 model we checked is under the MIT license, which broadly permits commercial use, modification and redistribution; keeping the copyright notice and license text is the main condition. DeepSeek-OCR-2 is Apache-2.0. Unlike Qwen, whose licenses change from model to model, the terms are consistent. This summary is not legal advice; read each model’s license text before use.

Running Them Locally

• The flagship models do not run on consumer GPUs. The memory requirements this site computed from the file sizes are about 570GB for V4.1-Flash and about 998GB for V4-Pro-0813. Even though only part of an MoE model is active per token, all the weights must be in memory. You need a server with several data-center GPUs or a machine with very large memory.
• There are no official GGUF builds. We cannot find any GGUF-format model on the deepseek-ai organization (as of 2026-09-26). Quantized builds come from third parties.
• There are small models too. DeepSeek-OCR-2 is small at about 3.4B, and its model card gives steps for running it with Transformers on NVIDIA GPUs. In June 2026 DeepSeek also released helper models for Qwen3 (4B, 8B, 14B) and Gemma 4 12B (with eagle3, dflash or dspark in their names), but they have no model cards, and their purpose and usage are not officially documented.
• Derived models such as uncensored versions are not the official models. The “Fine-tunes and Builds by Others" table below lists them separately from the articles on DeepSeek’s own models.

Sources: the deepseek-ai organization card; model cards of deepseek-ai/DeepSeek-V4-Flash, deepseek-ai/DeepSeek-V4-Flash-Vision-Exp, deepseek-ai/DeepSeek-V4.1-Flash and deepseek-ai/DeepSeek-OCR-2; the model list of the deepseek-ai organization on Hugging Face (all as of 2026-09-26). Memory requirements are from this site’s articles.

Our Coverage and Data

Local Model Watch has published 2 article(s) on DeepSeek’s own models, 2 on third-party fine-tunes and quantized builds based on them, and 0 from its official blog. The model list below is collected by code from Hugging Face every day. Part of our publisher index; model families link to their family pages.

At a Glance

Recent Models on Hugging Face

The publisher lists 105 models in total on Hugging Face. Below are up to 15 of the most recent generative models (text, image, video, audio and so on), the scope of this site, per organization. Classifiers, feature extractors, research components and repositories without a model card are omitted, and the publisher’s own quantized builds and helper variants are folded into the row of the original model. Parameter counts are left blank for repositories that pack weights into integer types, where Hugging Face’s count is not the real parameter count.

From the Hugging Face API. Last changed 2026-09-26 (JST). Collected by code.

Licenses of Its Own Models (as recorded in our articles)

The license can differ from model to model even within one publisher. This is not legal advice; check each model’s license text before use.

By Model Family

Articles on Its Own Models

Fine-tunes and Builds by Others

Last updated 2026-09-26 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log and the Hugging Face API.