DeepSeek AI: Models, Research and Public Weights

DeepSeek develops language and reasoning models, publishes checkpoints and technical reports, and offers hosted chat and API services.

kindCompany

Latest

Updated

Research artifacts and hosted services

DeepSeek develops models and distributes them through its chat application, API platform and public research channels. Its company site identifies the Hangzhou entity behind those services. The published checkpoints and reports let other developers examine work that would be hidden behind an API-only product.

Separate architecture, reasoning and cost claims

The V3 release used a mixture-of-experts architecture with 671 billion total parameters and 37 billion active per token. Its technical report’s roughly $5.6 million estimate concerns reported training compute, excluding prior research and ablations. It is not an audited estimate of the company’s full development cost.

R1 added a reasoning-training program and smaller distilled derivatives. The model card distinguishes R1 from R1-Zero, which explored reinforcement learning without the same preliminary supervised stage. Distilled models based on Qwen or Llama retain upstream licensing considerations.

Read access terms by checkpoint

DeepSeek’s October 2026 Flash update reached chat, app and API users and linked a downloadable checkpoint. API availability, a public weight file and the company’s own benchmark results establish different things. The release record supports those distribution paths without proving every performance claim or the economics of the whole lab.

Sources

  1. [1]
    DeepSeek | 深度求索

    DeepSeek · retrieved Oct 23, 2026

  2. [2]
    DeepSeek-V3 Technical Report

    arXiv · Dec 27, 2024

  3. [3]
    deepseek-ai/DeepSeek-R1

    DeepSeek · retrieved Oct 23, 2026

  4. [4]

DeepSeek history

2026

│
DeepSeek ships V4.1-Flash-1017 with open weights

A refresh of September's V4.1 Flash, served under the same deepseek-flash API name and published under the MIT license. DeepSeek's own table shows it ahead of its predecessor on every benchmark listed, for example 42.3 against 31.2 on Terminal-Bench 4.0, with shorter outputs that cut the cost of each task.

DeepSeek
Model
│
DeepSeek releases V4.1 Flash with open weights

DeepSeek released V4.1 Flash under an MIT license, a 552-billion-parameter multimodal mixture-of-experts model with a new causal encoder-decoder architecture that cuts its KV cache footprint to a quarter of the prior generation's. The company says it beats V4 Pro on performance, cost, and speed, and is routing all V4 Pro traffic to Flash within days.

DeepSeek
Model
│
Anthropic accuses seven Chinese labs of distilling Claude

With Anthropic, Alibaba, Moonshot AI, Xiaomi

Anthropic's September misuse report named Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax, and said Moonshot, DeepSeek and Xiaomi fed their own users' conversations into Claude. It also described a blocked request for help with gain-of-function work on chikungunya virus.

Anthropic
Research

2025

│
DeepSeek's math model scores 118 of 120 on the Putnam

DeepSeekMath-V2 trains a verifier to catch its own reasoning errors, then trains a proof generator against that verifier as the reward signal. The 685-billion-parameter model, released under Apache 2.0, scored 118 of 120 on the 2024 Putnam and gold-level on IMO 2025.

arXiv
Model
│
DeepSeek compresses text into pictures of text

DeepSeek-OCR treats a vision encoder as a compression scheme for context, packing a page into between 64 and 400 vision tokens depending on resolution. Released MIT licensed, and read less as an OCR model than as an argument about what context windows could cost.

DeepSeek
Model
■
DeepSeek open sources the infrastructure behind V3 and R1

Five repositories in five days, plus a bonus release: an MoE decoding kernel, a communication library, an FP8 GEMM library, pipeline-parallelism tools and a distributed file system, all described as production code. A closing report detailed the economics of running the models at scale.

DeepSeek
Open source
│
Microsoft and OpenAI investigate DeepSeek over possible distillation

With Microsoft, OpenAI

Microsoft's security team flagged accounts linked to DeepSeek pulling unusually large volumes of data through OpenAI's API in late 2024, and the companies opened a joint investigation into whether the outputs were used to train a rival model. OpenAI's terms bar using its outputs to build competing systems.

Tom's Hardware
Culture
│
DeepSeek releases Janus-Pro, an open text-to-image model

DeepSeek open sourced Janus-Pro, a unified model for image understanding and text-to-image generation in 1B and 7B sizes, a week after R1 upended the reasoning-model market. It claimed better prompt following and more stable generation than the original Janus.

DeepSeek
Model
●
DeepSeek releases R1 reasoning weights and smaller distilled models

DeepSeek released downloadable R1 reasoning-model weights and six smaller derivatives trained from its outputs. Its own tests put R1 near OpenAI o1 on several reasoning benchmarks, while the technical report documented a training approach others could study and limitations beyond those benchmarks.

DeepSeek
Model

2024

│
DeepSeek previews its first reasoning model, R1-Lite-Preview

The Chinese lab's new model shows its reasoning steps in real time and matches OpenAI's o1-preview on AIME and MATH benchmarks, with accuracy improving the longer it is allowed to think. It was free to test on DeepSeek's chat site, with open weights promised to follow.

DeepSeek
Model