DeepSeek Research Papers: Complete Guide to DeepSeek AI Research

Last updated: May 10, 2026

DeepSeek Research has become one of the most closely watched areas in open-weight AI because it connects large language model architecture, efficient training, reasoning, coding, math, multimodal systems, and long-context engineering in a fast-moving public research stream. This guide collects the most important DeepSeek Research papers, technical reports, model cards, repositories, and release notes so readers can understand what to read first, how the papers connect, and where to find official sources.

DeepSeek’s verified Hugging Face organization describes DeepSeek as a Chinese company founded in 2023 and dedicated to AGI research, while its official model pages and repositories host many of the papers, model cards, weights, and implementation notes discussed in this guide.

---

Quick Answer: What Are the Most Important DeepSeek Research Papers?

The most important DeepSeek papers to start with are DeepSeek LLM, DeepSeekMoE, DeepSeekMath, DeepSeek-Coder, DeepSeek-V2, DeepSeek-V3, DeepSeek-R1, DeepSeek-V3.2, and the official DeepSeek-V4 technical materials. Together, these sources explain DeepSeek’s path from scaling open language models to MoE architectures, Multi-head Latent Attention, GRPO-based reasoning, FP8 training, long-context efficiency, agentic coding, mathematical reasoning, theorem proving, multimodal research, and OCR-based context compression.

---

---

What Is DeepSeek Research?

DeepSeek Research refers to the body of AI research, technical reports, model releases, and open model infrastructure published by DeepSeek-AI. It covers general-purpose language models, mixture-of-experts models, reasoning models, code models, math models, theorem-proving systems, vision-language models, OCR/context-compression research, and AI infrastructure.

The research line began with open language and code models such as DeepSeek LLM and DeepSeek-Coder, then moved into efficiency-focused architectures such as DeepSeekMoE, DeepSeek-V2, DeepSeek-V3, and later DeepSeek-V3.2 and DeepSeek-V4. DeepSeek’s model cards and reports repeatedly emphasize efficient scaling, sparse computation, long-context inference, reinforcement learning, and open model access.

The phrase “DeepSeek Research papers” usually refers to official arXiv papers, DeepSeek technical reports, Hugging Face model cards, GitHub repositories, and peer-reviewed publications associated with DeepSeek-AI. Some sources are full research papers, while others are release notes or model cards. This distinction matters because a model card can document usage, weights, license, chat templates, or deployment details without being a full peer-reviewed paper.

---

Why DeepSeek Research Papers Matter

DeepSeek papers matter for five practical reasons.

First, they show how frontier-style language model development can be made more compute-efficient through sparse architectures. DeepSeekMoE proposed fine-grained expert segmentation and shared expert isolation to improve expert specialization in MoE language models. DeepSeek-V2 then applied DeepSeekMoE together with Multi-head Latent Attention to reduce inference costs and KV-cache requirements while scaling to 236B total parameters with 21B activated per token.

Second, DeepSeek’s reasoning papers helped popularize reinforcement-learning-centered reasoning research. DeepSeekMath introduced Group Relative Policy Optimization, or GRPO, in the context of mathematical reasoning, and DeepSeek-R1 later used large-scale reinforcement learning to incentivize reasoning behaviors. The R1 work was also published in Nature, where the authors describe using reinforcement learning to elicit reasoning patterns such as self-reflection and verification without relying on human-labeled reasoning trajectories.

Third, DeepSeek papers are useful for engineers because many releases include weights, model cards, GitHub repositories, and deployment notes. The DeepSeek-R1 repository, for example, provides R1, R1-Zero, and distilled models, while the DeepSeek-V3 and DeepSeek-V4 model cards document architecture, model sizes, context length, and local-running pathways.

Fourth, DeepSeek research connects multiple AI domains. The same research family includes code intelligence through DeepSeek-Coder and DeepSeek-Coder-V2, mathematical reasoning through DeepSeekMath and DeepSeekMath-V2, theorem proving through DeepSeek-Prover, multimodal systems through DeepSeek-VL, DeepSeek-VL2, Janus, and Janus-Pro, and long-context compression through DeepSeek-OCR.

Fifth, the research is timely. As of this update, DeepSeek’s official API documentation lists DeepSeek-V4-Pro and DeepSeek-V4-Flash as supported API models, with the legacy deepseek-chat and deepseek-reasoner names scheduled for discontinuation on July 24, 2026.

---

Complete List of Important DeepSeek Research Papers

The table below focuses on official or high-authority sources. Some entries are full arXiv papers; others are official release notes or Hugging Face/GitHub model cards. When an item is not a conventional paper, it is labeled accordingly.

Paper / Technical Report	Year	Main Topic	Model or System	Why It Matters	Official Source Type	Best For
DeepSeek LLM: Scaling Open-Source Language Models with Longtermism	2024	General LLM scaling	DeepSeek LLM 7B/67B	Establishes the early DeepSeek LLM line, scaling-law study, SFT/DPO chat models, and 2T-token pretraining dataset.	arXiv + GitHub	Beginners, LLM foundations
DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models	2024	MoE architecture	DeepSeekMoE	Introduces fine-grained expert segmentation and shared expert isolation, which become central to later DeepSeek MoE models.	arXiv	Architecture researchers
DeepSeek-Coder	2024	Code LLMs	DeepSeek-Coder 1.3B–33B	Introduces open code models trained on project-level code data, with code generation and infilling focus.	arXiv	Software engineers, coding agents
DeepSeekMath	2024	Math reasoning	DeepSeekMath 7B	Introduces a math-focused model and GRPO, a key method later associated with DeepSeek reasoning work.	arXiv + GitHub	Reasoning researchers, students
DeepSeek-VL	2024	Vision-language understanding	DeepSeek-VL	Applies DeepSeek research to real-world multimodal tasks such as screenshots, PDFs, OCR, charts, and scientific documents.	arXiv + GitHub	Multimodal engineers
DeepSeek-V2	2024	Efficient MoE LLM	DeepSeek-V2	Combines DeepSeekMoE and Multi-head Latent Attention; supports 128K context and activates 21B of 236B parameters per token.	arXiv + GitHub	LLM infra teams
DeepSeek-Prover	2024	Formal theorem proving	DeepSeek-Prover	Uses synthetic Lean 4 theorem-proving data to improve formal proof generation.	arXiv	Formal math researchers
DeepSeek-Coder-V2	2024	MoE code intelligence	DeepSeek-Coder-V2	Continues from DeepSeek-V2 with code-focused training, 128K context, and support for hundreds of programming languages.	arXiv + GitHub	Coding model builders
DeepSeek-Prover-V1.5	2024 / 2025	Lean theorem proving	DeepSeek-Prover-V1.5	Uses proof assistant feedback, reinforcement learning, and Monte Carlo tree search; published as an ICLR 2025 conference paper.	arXiv + ICLR	Theorem proving
Janus	2024	Unified multimodal understanding and generation	Janus	Decouples visual encoding for multimodal understanding and generation in a unified autoregressive framework.	arXiv + GitHub	Multimodal research
JanusFlow	2024 / 2025	Multimodal generation	JanusFlow	Combines autoregressive modeling with rectified flow; accepted in CVPR 2025 open access proceedings.	arXiv + CVPR	Vision-generation researchers
DeepSeek-VL2	2024	MoE vision-language model	DeepSeek-VL2	Adds dynamic tiling and DeepSeekMoE/MLA-based language modeling for advanced multimodal understanding.	arXiv + GitHub	VLM builders
DeepSeek-V3 Technical Report	2024 / 2025	Efficient frontier-scale MoE LLM	DeepSeek-V3	Introduces a 671B-parameter MoE with 37B activated per token, MLA, DeepSeekMoE, auxiliary-loss-free balancing, multi-token prediction, and FP8 training.	arXiv + GitHub	LLM architecture and training teams
DeepSeek-R1	2025	Reasoning via RL	DeepSeek-R1, R1-Zero, distilled models	A major reasoning-model release that studies RL-driven reasoning and distillation into smaller dense models.	arXiv + GitHub + Nature	Reasoning researchers
Janus-Pro	2025	Multimodal understanding and generation	Janus-Pro	Scales and improves Janus with better training strategy, data, and model size.	arXiv + Hugging Face/GitHub	Multimodal product teams
DeepSeek-R1-0528	2025	R1 update	DeepSeek-R1-0528	Official R1 update with improved benchmarks, JSON output, function calling, and open weights.	DeepSeek API docs + Hugging Face	API users, app developers
DeepSeek-V3.1 / V3.1-Terminus	2025	Hybrid thinking/non-thinking and agents	DeepSeek-V3.1	Official release notes describe hybrid inference, faster thinking, and improved agent capabilities; Terminus addresses language consistency and agent feedback.	DeepSeek API docs + Hugging Face	Agent builders
DeepSeek-V3.2-Exp	2025	Sparse attention experiment	DeepSeek-V3.2-Exp	Introduces DeepSeek Sparse Attention as an experimental step toward more efficient long-context training and inference.	DeepSeek API docs + GitHub	Long-context researchers
DeepSeek-OCR	2025	Optical context compression	DeepSeek-OCR	Studies visual-text compression for long-context processing using DeepEncoder and a DeepSeek3B-MoE decoder.	arXiv + GitHub	Document AI, OCR, context compression
DeepSeekMath-V2	2025	Self-verifiable math reasoning	DeepSeekMath-V2	Investigates LLM-based verification and proof generation for rigorous mathematical reasoning.	arXiv	Advanced math reasoning
DeepSeek-V3.2	2025	Reasoning, DSA, agents	DeepSeek-V3.2 / V3.2-Speciale	Adds DSA, scalable RL, and agentic task synthesis; official sources describe it as reasoning-first and built for agents.	arXiv + DeepSeek docs + Hugging Face	Agentic AI and reasoning teams
DeepSeek-V4 Technical Materials	2026	Million-token context, MoE, hybrid attention	DeepSeek-V4-Pro / V4-Flash	Official V4 sources describe 1M context, V4-Pro at 1.6T total / 49B active parameters, V4-Flash at 284B total / 13B active, CSA/HCA hybrid attention, mHC, and Muon optimizer.	DeepSeek docs + Hugging Face technical report/model card	Long-context AI, infra, agents

---

Best Papers to Start With

Start here if you are new to DeepSeek Research:

1. DeepSeek LLM for the early model family and scaling philosophy.
2. DeepSeekMoE for the sparse expert architecture.
3. DeepSeek-V2 for MLA and efficient inference.
4. DeepSeek-V3 for large-scale MoE training and FP8-oriented efficiency.
5. DeepSeek-R1 for reinforcement-learning-based reasoning.
6. DeepSeek-V3.2 for sparse attention, agentic training, and reasoning-first design.
7. DeepSeek-V4 for the current million-token context direction.

---

Recommended Reading Order for DeepSeek Research Papers

Reader Type	Start With	Then Read	Why This Order Works
Beginner	DeepSeek LLM	DeepSeek-V2, DeepSeek-V3, DeepSeek-R1	Moves from general LLM scaling to efficient MoE models and reasoning.
LLM architecture researcher	DeepSeekMoE	DeepSeek-V2, DeepSeek-V3, DeepSeek-V3.2, DeepSeek-V4	Follows the architecture path from expert specialization to MLA, DSA, and V4 hybrid attention.
Reasoning/RL researcher	DeepSeekMath	DeepSeek-R1, DeepSeek-R1-0528, DeepSeek-V3.2, DeepSeekMath-V2	Shows the progression from GRPO to large-scale RL, distillation, and verification-based reasoning.
Coding model engineer	DeepSeek-Coder	DeepSeek-Coder-V2, DeepSeek-R1, DeepSeek-V3.2, DeepSeek-V4	Connects code pretraining, long context, reasoning, tool use, and agentic coding.
Math/theorem proving researcher	DeepSeekMath	DeepSeek-Prover, DeepSeek-Prover-V1.5, DeepSeekMath-V2	Moves from informal math reasoning to Lean proof generation and self-verification.
Multimodal/OCR researcher	DeepSeek-VL	DeepSeek-VL2, Janus, Janus-Pro, DeepSeek-OCR	Covers vision-language understanding, unified generation, and visual context compression.
Infrastructure team	DeepSeek-V2	DeepSeek-V3, V3.2-Exp, V3.2, V4, DeepSeek GitHub infrastructure repos	Focuses on model architecture, efficient attention, context length, kernels, and deployment.

---

DeepSeek Research Timeline

2023: DeepSeek begins releasing open model work, including early LLM and code-model repositories. The official DeepSeek LLM repository describes the 7B/67B model family and the 2T-token bilingual pretraining dataset.

January 2024: DeepSeek LLM and DeepSeekMoE appear on arXiv. DeepSeek LLM focuses on scaling open language models, while DeepSeekMoE focuses on expert specialization in mixture-of-experts language models.

January–February 2024: DeepSeek-Coder and DeepSeekMath are published. DeepSeek-Coder focuses on code intelligence, while DeepSeekMath introduces mathematical reasoning work and GRPO.

March 2024: DeepSeek-VL is released as a vision-language model for real-world multimodal understanding, including screenshots, PDFs, OCR, charts, and scientific documents.

May–June 2024: DeepSeek-V2 and DeepSeek-Coder-V2 extend the research line into efficient MoE inference and code-focused continued training.

August 2024: DeepSeek-Prover-V1.5 appears as an improvement over DeepSeek-Prover, using proof assistant feedback, reinforcement learning, and Monte Carlo tree search. It later appears as an ICLR 2025 conference paper.

December 2024: DeepSeek-V3 is released as a major MoE technical report with 671B total parameters, 37B activated per token, MLA, DeepSeekMoE, auxiliary-loss-free load balancing, multi-token prediction, and large-scale pretraining.

January 2025: DeepSeek-R1 is released, including R1-Zero, R1, and distilled dense models. The work centers on reinforcement learning for reasoning and distillation into smaller models.

January 2025: Janus-Pro is released as an improved Janus model for unified multimodal understanding and generation.

May 2025: DeepSeek-R1-0528 is released with official notes describing improved benchmark performance, reduced hallucinations, JSON output, function calling, and open weights.

September 2025: DeepSeek-V3.2-Exp is released as an experimental model built on V3.1-Terminus and introducing DeepSeek Sparse Attention for long-context efficiency research.

October–November 2025: DeepSeek-OCR and DeepSeekMath-V2 expand the research line into optical context compression and self-verifiable mathematical reasoning.

December 2025: DeepSeek-V3.2 and V3.2-Speciale are released. Official sources describe V3.2 as a reasoning-first model built for agents, with DSA, scalable RL, and a large-scale agentic task synthesis pipeline.

April 2026: DeepSeek-V4 Preview is released with V4-Pro and V4-Flash, official 1M context support, API availability, open weights, and technical materials on Hugging Face.

---

Key Technical Themes in DeepSeek Research

Mixture-of-Experts and DeepSeekMoE

Mixture-of-Experts, or MoE, is an architecture where only part of a large model is activated for each token. The goal is to increase total model capacity without paying the full compute cost of activating every parameter. DeepSeekMoE introduced a specialization-focused MoE design using fine-grained experts and shared experts. Later models such as DeepSeek-V2, DeepSeek-V3, V3.2, and V4 build heavily on sparse expert design.

Multi-head Latent Attention

Multi-head Latent Attention, or MLA, is one of DeepSeek’s central efficiency ideas. In plain English, MLA compresses the attention key-value cache into a smaller latent representation, which can reduce memory pressure during inference. DeepSeek-V2 introduced MLA together with DeepSeekMoE, and DeepSeek-V3 further validated that architecture at larger scale.

Efficient Training and FP8

DeepSeek-V3 is the key paper for efficient large-scale training. The report describes a 671B-parameter MoE model with 37B activated parameters per token and reports pretraining on 14.8T tokens. It also highlights efficiency-focused choices such as MLA, DeepSeekMoE, auxiliary-loss-free balancing, multi-token prediction, and stable large-scale training.

Reinforcement Learning and GRPO

GRPO appears in the DeepSeekMath paper as a reinforcement learning method for mathematical reasoning. DeepSeek-R1 later makes RL the center of the reasoning pipeline, with R1-Zero exploring reasoning behavior through RL before the more refined R1 pipeline adds cold-start data and multi-stage training. The Nature version of the R1 paper frames this as a demonstration that reasoning can be incentivized through reinforcement learning without human-labeled reasoning trajectories.

Reasoning Models and Distillation

DeepSeek-R1 is important not only because of its own reasoning capability, but because it shows how reasoning traces from a large model can be distilled into smaller dense models. The official R1 repository lists distilled checkpoints based on Qwen and Llama model families, giving researchers a way to study smaller reasoning models without running the full 671B-parameter MoE model.

Long-Context and Sparse Attention

DeepSeek’s long-context work evolves from MLA in V2 and V3 to DeepSeek Sparse Attention in V3.2-Exp and V3.2, then to V4’s million-token context technical materials. V3.2’s arXiv abstract identifies DSA as a key mechanism for reducing computational complexity while preserving long-context performance, while V4’s official model card describes a hybrid attention design that combines Compressed Sparse Attention and Heavily Compressed Attention.

Coding and Agentic Workflows

DeepSeek-Coder and DeepSeek-Coder-V2 provide the code-model foundation. DeepSeek-V3.1, V3.2, and V4 then move toward agentic use cases such as tool use, code agents, and long-horizon task execution. Official V3.2 sources describe “thinking in tool-use,” and V4 release notes emphasize agentic coding and integration with leading AI-agent workflows.

Mathematical Reasoning and Proof Generation

DeepSeekMath introduced a math-focused model and GRPO. DeepSeek-Prover and DeepSeek-Prover-V1.5 then extended this direction into Lean 4 theorem proving, synthetic formal data, proof assistant feedback, and tree search. DeepSeekMath-V2 continues the line by focusing on self-verifiable mathematical reasoning and LLM-based verification.

Multimodal, Vision-Language, and OCR Research

DeepSeek-VL and DeepSeek-VL2 focus on vision-language understanding. Janus, JanusFlow, and Janus-Pro address unified multimodal understanding and generation. DeepSeek-OCR takes a different route by studying whether visual representations can compress long text contexts through optical 2D mapping.

---

Technical Themes by Paper

Technical Theme	Best DeepSeek Source	What It Explains	Practical Value
Scaling open LLMs	DeepSeek LLM	Scaling laws, 7B/67B models, SFT/DPO chat models	Foundation for understanding the early DeepSeek stack
Expert specialization	DeepSeekMoE	Fine-grained expert segmentation and shared experts	Efficient model capacity expansion
KV-cache efficiency	DeepSeek-V2	MLA and MoE for cheaper inference	Long-context serving and lower memory pressure
FP8 and large MoE training	DeepSeek-V3	Stable training, MoE scale, multi-token prediction	Frontier-scale training design
RL for reasoning	DeepSeekMath, DeepSeek-R1	GRPO, RL, R1-Zero, R1, distillation	Math, coding, STEM, and chain-of-reasoning research
Tool use and agents	DeepSeek-V3.2	Thinking in tool use and agentic task synthesis	Agent workflows, code agents, tool-using models
Sparse attention	V3.2-Exp, V3.2	DeepSeek Sparse Attention	Long-context efficiency
Million-token context	DeepSeek-V4	CSA/HCA hybrid attention, 1M context	Large-document, repository, and agent memory use cases
Formal proof generation	DeepSeek-Prover, Prover-V1.5	Lean 4, synthetic proof data, RL from proof assistant feedback	Formal math and theorem proving
Vision-language AI	DeepSeek-VL, VL2, Janus-Pro	OCR, document understanding, multimodal generation	Multimodal products
Optical compression	DeepSeek-OCR	Vision-text compression for long contexts	Document AI and context compression

---

DeepSeek Research Papers by Use Case

If You Want To…	Read First	Then Read	Why
Understand DeepSeek reasoning models	DeepSeekMath	DeepSeek-R1, Nature R1 paper, DeepSeekMath-V2	Shows GRPO, RL-based reasoning, and self-verification.
Build coding agents	DeepSeek-Coder	DeepSeek-Coder-V2, DeepSeek-R1, V3.2, V4	Moves from code pretraining to agentic reasoning and tool use.
Study efficient LLM training	DeepSeekMoE	DeepSeek-V2, DeepSeek-V3, V4	Explains sparse experts, MLA, FP8-oriented training, and hybrid attention.
Study long context	DeepSeek-V2	V3.2-Exp, V3.2, V4, DeepSeek-OCR	Covers MLA, DSA, 1M context, and visual context compression.
Study math reasoning	DeepSeekMath	DeepSeek-R1, DeepSeekMath-V2	Follows math RL from answer-based reasoning to verification.
Study theorem proving	DeepSeek-Prover	Prover-V1.5, DeepSeekMath-V2	Focuses on Lean, proof assistant feedback, and proof verification.
Build multimodal apps	DeepSeek-VL	DeepSeek-VL2, Janus-Pro, DeepSeek-OCR	Covers document understanding, OCR, image understanding, and generation.
Run models locally	GitHub + Hugging Face model cards	V3, R1, V3.2, V4 model cards	Model cards explain weights, chat templates, supported runtimes, and license details.

---

Official Sources for DeepSeek Research Papers

Use these source categories when verifying DeepSeek Research papers:

1. DeepSeek official website and API docs
DeepSeek release notes and API documentation are the best source for current model availability, API model names, deprecations, context length, and release dates. The V4 release page, for example, lists V4-Pro and V4-Flash, official 1M context support, open weights, and the V4 technical report link.

2. DeepSeek GitHub repositories
GitHub repositories often include paper links, quick-start instructions, model downloads, licensing notes, and citations. They are especially useful for DeepSeek LLM, V2, V3, R1, Coder-V2, VL, VL2, Janus, and OCR.

3. DeepSeek verified Hugging Face organization
Hugging Face is important for official model cards, collections, weights, downloads, inference examples, and technical summaries. The verified organization lists collections such as DeepSeek-V4 and DeepSeek-OCR.

4. arXiv
arXiv is the main source for most DeepSeek technical papers, including DeepSeek LLM, DeepSeekMoE, DeepSeekMath, DeepSeek-V2, DeepSeek-Coder-V2, DeepSeek-V3, DeepSeek-R1, DeepSeek-V3.2, DeepSeek-OCR, and DeepSeekMath-V2.

5. Peer-reviewed or conference sources
Some DeepSeek-related research appears in peer-reviewed or conference venues. DeepSeek-R1 has a Nature article, DeepSeek-Prover-V1.5 appears as an ICLR 2025 conference paper, and JanusFlow appears in CVPR 2025 open access proceedings.

Official Links Checklist

Before citing or using a DeepSeek model, verify:

• Is the source from deepseek-ai on GitHub or Hugging Face?
• Is there an arXiv paper or official technical report?
• Is the release note on official DeepSeek API docs?
• Does the model card list the license and usage restrictions?
• Is the benchmark self-reported, third-party, or peer-reviewed?
• Is the model a full release, preview, experiment, distilled model, or community quantization?
• Does the source include the current chat template, context length, and deployment instructions?

---

Common Misconceptions About DeepSeek Research

DeepSeek Research vs. the DeepSeek Chatbot

DeepSeek Research is broader than the chatbot. The chatbot is an application interface for certain models, while the research includes papers, model weights, repositories, model cards, API releases, infrastructure libraries, and training/inference techniques.

Papers vs. Model Cards vs. GitHub Repositories

A paper explains the research method. A model card explains the released model, license, usage, limitations, and deployment guidance. A GitHub repository may include code, model links, examples, and citations. For DeepSeek, all three can be important.

Open Weight vs. Fully Open Source

A model can have public weights without every part of the training pipeline being open. Always check the license, code license, model license, data disclosure, and intended-use notes. DeepSeek often publishes weights and code, but not every release has the same level of openness or documentation.

Benchmarks vs. Real-World Performance

DeepSeek papers and model cards include many benchmark claims, but benchmark performance is not the same as production reliability. Treat official benchmark tables as vendor-reported unless a third-party or peer-reviewed source independently verifies them. This is especially important for comparisons against closed-source models.

DeepSeek-V4 Is Not Just “A Bigger V3”

DeepSeek-V4 official materials describe new architectural and optimization ideas, including hybrid CSA/HCA attention, mHC, Muon optimizer, 1M context, and two model sizes. It should be read as a new technical direction, not merely as a larger checkpoint.

---

How to Cite DeepSeek Research Papers

For academic work, cite the latest official arXiv version, conference version, or journal version. Do not fabricate BibTeX. Use the citation block from the official arXiv page, GitHub repository, Hugging Face model card, Nature article, ICLR page, or CVPR page.

General guidance:

• For arXiv papers: cite the arXiv ID and version you used.
• For GitHub repositories: cite the repository, commit hash if relevant, and access date.
• For Hugging Face model cards: cite the model name, organization, revision if relevant, and access date.
• For API release notes: cite the official DeepSeek API documentation page and access date.
• For peer-reviewed work: cite the published Nature, ICLR, or CVPR version when available.

Example citation formats:

DeepSeek-AI. “DeepSeek-V3 Technical Report.” arXiv:2412.19437, 2024/2025.

DeepSeek-AI. “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.” arXiv:2501.12948, 2025.

DeepSeek-AI. “DeepSeek-V4-Pro.” Hugging Face model card, accessed May 10, 2026.

DeepSeek-AI. “DeepSeek V4 Preview Release.” DeepSeek API Docs, April 24, 2026.

---

FAQs About DeepSeek Research Papers

---

Conclusion

DeepSeek Research is best understood as a connected research program rather than a single paper or model. The core path starts with DeepSeek LLM, moves through DeepSeekMoE, DeepSeek-V2, and DeepSeek-V3, then branches into DeepSeek-R1 for reasoning, DeepSeek-Coder for programming, DeepSeekMath and DeepSeek-Prover for mathematics, DeepSeek-VL/Janus for multimodal systems, DeepSeek-OCR for visual context compression, and DeepSeek-V3.2/V4 for long-context and agentic AI.

For most readers, the best reading order is: DeepSeek LLM → DeepSeekMoE → DeepSeek-V2 → DeepSeek-V3 → DeepSeek-R1 → DeepSeek-V3.2 → DeepSeek-V4. After that, follow the specialized track that matches your work: coding, math, theorem proving, multimodal AI, OCR, or infrastructure.

Bookmark this guide and revisit it whenever DeepSeek publishes a new model, paper, model card, or technical report.