A Guide to DeepSeek Research Papers From V2 to V4

Map every major DeepSeek research paper from V2 to the V4 technical report, with key contributions and benchmarks. Start reading today.

News·April 24, 2026·By DS Guide Editorial
DeepSeek research papers title card: eyebrow 'News · Research papers', title 'Research Papers', tagline 'From DeepSeek LLM scaling law to the V4 technical report.', and four stats: Jan 2024, first paper; arXiv, where to read; Apr 2026, V4 report; Open, weights + code; beside an open-book illustration.

If you are trying to understand why the DeepSeek-V4-Pro preview needs only 10% of the KV cache of V3.2 at a one-million-token context, or how a Chinese lab trained a reasoning model for under $300,000, the answer sits inside the DeepSeek research papers — not the press coverage. This guide maps every major publication the lab has put on arXiv, from the original DeepSeek LLM scaling-law work in January 2024 to the DeepSeek-V4 technical report, posted to arXiv on 26 April 2026, and its September 2026 V4.1-Flash follow-up. For each paper we note the model it describes, the single architectural idea it contributed, and how those ideas stack into the V4 release.

You will finish with a reading order, a benchmark cross-reference, and pointers to our deeper model-by-model breakdowns.

Why DeepSeek’s papers matter more than most

Illustration: a reader studying a research-paper PDF on a laptop at a desk, with a printed copy, highlighted passages and handwritten notes

Most frontier AI labs ship products and a blog post. DeepSeek ships products, open weights, and a detailed technical report — usually on arXiv, usually with enough implementation detail to reproduce the architecture. That is how the lab became a reference point in 2025: DeepSeek drew wide attention in 2025 when its R1 release delivered near industry-leading reasoning performance – for allegedly a fraction of the price.

The V4 preview, released on April 24, 2026, continues that pattern. Its paper (arXiv, 26 April 2026) presents the April 2026 preview versions of two Mixture-of-Experts language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. Those sizes describe the preview models, not the DeepSeek-V4-Pro-0813 and DeepSeek-V4.1-Flash-1017 models the API serves today. If you want to understand what V4 actually does differently, you read the V4 technical report. This article is a reading guide, not a replacement for those primary sources.

The DeepSeek paper timeline at a glance

Every paper below is first-party DeepSeek-AI work. We have limited the list to the releases that matter for understanding the V4 generation. For a wider bibliography, the community-maintained DeepSeek Papers collection on Hugging Face is the most complete single index.

Paper Published arXiv / host Core contribution
DeepSeek LLM 2024-01-05 arXiv 2401.02954 Scaling-law study for 7B and 67B dense models
DeepSeekMoE 2024-01-11 arXiv 2401.06066 Fine-grained expert specialisation
DeepSeek-Coder 2024-01-25 arXiv 2401.14196 Code-specialised base models
DeepSeekMath (GRPO) 2024-02-05 arXiv 2402.03300 Group Relative Policy Optimisation
DeepSeek-V2 2024-05-07 arXiv 2405.04434 Multi-head Latent Attention (MLA)
DeepSeek-V3 2024-12-27 arXiv 2412.19437 Auxiliary-loss-free load balancing, FP8 training
DeepSeek-R1 2025-01-22 arXiv 2501.12948 RL-only reasoning training
DeepSeek-V3.2 2025-12-02 arXiv 2512.02556 DeepSeek Sparse Attention (DSA)
DeepSeek-V4 (preview) 2026-04-26 arXiv 2606.19348 Hybrid CSA + HCA attention, 1M-token context
DeepSeek-V4.1-Flash 2026-09-17 arXiv 2609.19969 KV cache compression (CSA2 with FP4 KV caching)

The foundation papers (2024)

DeepSeek LLM — the scaling-law study

The first DeepSeek paper is not about a frontier model. It is a scaling-law study that justifies everything that came after. The authors argue that prior scaling literature gave conflicting guidance, so they re-derived the laws for their own training stack and applied them to 7B and 67B dense configurations. The published corpus at the time was 2 trillion tokens and growing. If you want context on why later DeepSeek models look the way they do — and why the lab keeps publishing token-count figures prominently — this is the origin. For a plain-English summary, see our page on DeepSeek LLM.

DeepSeekMoE and DeepSeek-V2 — the architecture base

DeepSeekMoE (January 2024) introduced the fine-grained expert split that every later DeepSeek MoE model inherits. DeepSeek-V2 (May 2024) then combined that MoE design with Multi-head Latent Attention (MLA), which is the attention trick that makes long-context serving affordable. DeepSeek-V2 comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times.

Every subsequent paper assumes you know what MLA and fine-grained MoE are. If you read only one paper from 2024, read this one. Our deeper breakdowns sit on the DeepSeek MoE and DeepSeek V2 pages.

DeepSeek-Coder and DeepSeekMath

Two specialist papers from early 2024 matter for the V4 story. DeepSeek-Coder described a series of open code models from 1.3B to 33B parameters trained from scratch on 2 trillion tokens. DeepSeekMath 7B continued pre-training from DeepSeek-Coder-Base-v1.5 7B and introduced Group Relative Policy Optimisation (GRPO), the RL algorithm that R1 would later use to learn reasoning without human demonstrations. Readers focused on coding should also see our notes on DeepSeek Coder and the DeepSeek Math page.

DeepSeek V3 — the efficiency thesis in paper form

The V3 technical report is the one most cited in the January 2025 news cycle. The paper presents DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for load balancing and sets a multi-token prediction training objective for stronger performance. DeepSeek-V3 was pre-trained on 14.8 trillion diverse and high-quality tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages.

The paper’s most-quoted number is the training cost: “Combined with 119K GPU hours for the context length extension and 5K GPU hours for post-training, DeepSeek-V3 costs only 2.788M GPU hours for its full training. Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M.” That is the figure widely reported as “about $5.6M.” The paper is careful to note this excludes prior research and ablations — a caveat many press summaries drop. Our DeepSeek V3 review covers the model in more depth.

DeepSeek R1 — reasoning without human demonstrations

R1 is the paper that turned DeepSeek from “efficient lab” to “strategic concern” in Western press coverage. The core claim is that reasoning steps can be taught with pure RL, no supervised step-by-step reasoning data required. The reasoning abilities of LLMs can be incentivized through pure reinforcement learning (RL), obviating the need for human-labeled reasoning trajectories. The proposed RL framework facilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification, and dynamic strategy adaptation. The trained model achieves superior performance on verifiable tasks such as mathematics, coding competitions, and STEM fields.

Two points readers often get wrong:

  • The training cost figure belongs to R1, not V3. Reuters reported DeepSeek’s own R1 training estimate at $294,000 in September 2025, and the revised R1 paper on arXiv (v2, 4 January 2026) tabulates it as 147K H800 GPU hours, or $294K at an assumed $2 per GPU hour. The $5.576M figure in the V3 paper is a separate model.
  • R1 builds on V3-Base. The paper uses DeepSeek-V3-Base as the base model, DeepSeek-V3 as the instructed model. DeepSeek-R1 and DeepSeek-R1-Zero are trained on top of DeepSeek-V3-Base and DeepSeek-R1 leverages non-reasoning data from DeepSeek-V3 SFT data. The two papers are a pair.

Our DeepSeek R1 review and the DeepSeek R1 vs OpenAI o1 comparison pair well with the paper.

DeepSeek V3.2 — the sparse attention bridge

Illustration: schematic chart of benchmark score against active parameters, with V3 marked high at low active parameters (not data from the paper)

V3.2 is the under-rated paper. Released in December 2025, it is the one that lays the groundwork for V4’s efficiency claims. The paper introduces DeepSeek Sparse Attention (DSA), an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. By implementing a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2 performs comparably to GPT-5. Notably, the high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI).

V3.2’s DSA is instantiated on top of MLA, not as a replacement for it. The V3.2 paper is the most detailed account of how DeepSeek thinks about attention efficiency before the bigger leap in V4. The DeepSeek V3.2 page summarises the release notes.

DeepSeek V4 — the April 2026 preview paper

The V4 technical report, titled DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence, was posted to arXiv on 26 April 2026 (arXiv 2606.19348) and is linked from the V4-Pro model card on Hugging Face. It describes the April 2026 preview models and is the primary source for every V4 claim below. DeepSeek has since posted a follow-up, DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression (arXiv 2609.19969, 17 September 2026).

Architecture contributions

The paper’s headline contribution is a new hybrid attention stack that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to improve long-context efficiency. In the one-million-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. The paper also introduces Manifold-Constrained Hyper-Connections (mHC), which constrain the residual mapping to enhance the stability of signal propagation across layers while preserving model expressivity, and uses the Muon optimizer.

Two more items are worth extracting from the report:

  • Precision. The routed expert parameters use FP4 precision, added through FP4 quantization-aware training in post-training; the April preview model card lists the released checkpoints as “FP4 + FP8 Mixed” (MoE expert parameters in FP4, most other parameters in FP8).
  • Three reasoning-effort modes. The paper’s DeepSeek-V4-Pro and DeepSeek-V4-Flash both support three modes: Non-think, Think High and Think Max. On today’s API, thinking is on by default, reasoning_effort takes "low", "high" or "max", and non-thinking mode is selected with {"thinking": {"type": "disabled"}}. In thinking mode the API returns reasoning_content alongside the final content.

Positioning against closed frontier models

The report is unusually candid about where V4 sits: “Through the expansion of reasoning tokens, DeepSeek-V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performance falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months.” That is a self-description the press did not have to spin.

In the paper’s comparison table (DeepSeek-reported, April 2026, against Opus-4.6, GPT-5.4, Gemini-3.1-Pro, K2.6 and GLM-5.1), DeepSeek-V4-Pro-Max posts the top scores on LiveCodeBench (93.5), Codeforces rating (3206) and Apex Shortlist (90.2); the paper adds that it ranks 23rd among human candidates on the Codeforces leaderboard. On SimpleQA-Verified its 57.9 leads the open models in the table but trails Gemini-3.1-Pro (75.6). On Putnam-2025, under a hybrid formal-informal regime, DeepSeek-V4 reaches a proof-perfect 120/120. These figures describe the April preview model; verify any individual number against the paper before you quote it — frontier benchmark standings shift monthly.

How the papers connect: a suggested reading order

If you have a weekend and want the architecture thread, read in this sequence:

  1. DeepSeek-V2 for MLA and DeepSeekMoE — the foundation.
  2. DeepSeek-V3 for auxiliary-loss-free load balancing, multi-token prediction, and FP8 training.
  3. DeepSeek-R1 for the RL-only reasoning recipe (which uses GRPO from DeepSeekMath).
  4. DeepSeek-V3.2 for DSA as an attention-efficiency prototype.
  5. DeepSeek-V4 for the CSA + HCA hybrid and the full mHC / Muon / FP4 training stack.

If you only have an evening, read the V4 report and V3 report back to back. Together they cover most of the V4 generation’s core ideas. The DeepSeek benchmarks 2026 page has the cross-model numbers.

Using the papers as a developer, not a reader

DeepSeek’s current models are served through POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com. As of November 2026 the API lists two model IDs: deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813) — later releases than the preview models in the V4 paper. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted but are served by DeepSeek-V4.1-Flash-1017; deepseek-chat and deepseek-reasoner were scheduled for retirement on 2026-07-24 15:59 UTC and are no longer listed in the official docs.

A minimal Python call against V4-Pro, matching the paper’s thinking-mode setup, looks like this:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.deepseek.com",
    api_key="YOUR_KEY",
)

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": "Summarise the V4 paper's hybrid attention section."}],
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)
print(resp.choices[0].message.reasoning_content)
print(resp.choices[0].message.content)

A few practical details for anyone replicating paper results against the hosted API:

  • The API is stateless. Resend the full messages array with every call; DeepSeek does not keep prior turns for you.
  • Context window is 1,000,000 tokens on both current API models, with output up to 384,000 tokens.
  • Temperature guidance from DeepSeek’s docs (non-thinking mode only): 0.0 for code and maths, 1.0 for data tasks, 1.3 for general chat and translation, 1.5 for creative writing. Thinking mode, the default, accepts temperature but ignores it.
  • JSON mode is designed to return valid JSON, not guaranteed. Prompt explicitly with the word “json” plus a small schema example, and set max_tokens high enough to avoid truncation.

For current per-token rates, peak and off-peak, see our DeepSeek API pricing page.

For the full API surface, start with the DeepSeek API documentation and the DeepSeek API getting started tutorial.

What the papers do not tell you

DeepSeek research papers lineage title card: eyebrow 'News · Research papers', title 'Paper Lineage', tagline 'Architecture papers on arXiv, May 2024 to April 2026.', and four stats: V2, MLA, May 2024; V3, MTP, Dec 2024; V3.2, DSA, Dec 2025; V4, CSA+HCA, Apr 2026; beside an open-manual illustration.

Three honest limitations of reading DeepSeek’s papers as your only source:

  • Self-reported benchmarks are self-reported. The V4 report’s evaluation configurations are the paper’s own. Cross-reference against independent benchmark aggregators before making a procurement decision.
  • Training-cost figures exclude prior research. The V3 paper says this explicitly. The $5.576M number is the official run, not the full R&D bill.
  • Licensing is per-model. Every recent DeepSeek release (V4.1-Flash-1017, V4-Pro-0813, V4.1-Flash, V3.2, V3.1, R1) publishes both code and model weights under MIT. Some older releases — V3 base, Coder-V2, VL2 — split code (MIT) from weights (a separate DeepSeek Model License). Check the specific Hugging Face repo if licensing matters to your deployment.

For broader independent context, the DeepSeek latest updates feed tracks post-publication corrections and community replications.

Frequently asked questions

Where can I find the official DeepSeek research papers?

DeepSeek publishes technical reports on arXiv and links them from the matching Hugging Face model cards. V4.1-Flash is arXiv 2609.19969; V4 is arXiv 2606.19348; V3 is arXiv 2412.19437; V3.2 is arXiv 2512.02556; R1 is arXiv 2501.12948; V2 is arXiv 2405.04434. A community-maintained index collects the older papers too. For our plain-English summaries of each release, see the DeepSeek models hub.

What does the DeepSeek V4 paper actually introduce?

The V4 technical report introduces a hybrid attention stack — Compressed Sparse Attention plus Heavily Compressed Attention — that the authors claim cuts single-token inference FLOPs to 27% and KV cache to 10% of V3.2 at 1M-token context. It also documents mHC residual connections, Muon-optimised training, and FP4 expert precision. Full details are in the paper (arXiv 2606.19348); release context is on our DeepSeek V4 release date page.

How much did it cost to train DeepSeek V3 and R1?

The V3 technical report reports 2.788M H800 GPU hours at an assumed $2/hour, totalling about $5.576M — explicitly excluding prior research and ablations. DeepSeek’s separately disclosed R1 training cost was around $294,000 (Reuters, September 2025; also tabulated in the revised R1 paper, arXiv v2 of 4 January 2026). The two figures belong to different models and should not be combined. Our DeepSeek V3 review unpacks the cost claim.

Is the DeepSeek R1 paper really a reasoning trace breakthrough?

The R1 paper’s contribution is showing that reasoning patterns like self-reflection and verification can emerge from pure reinforcement learning, without human-labelled step-by-step reasoning data. That is a methodological result, not a claim to the best possible reasoning model. Compare it honestly against closed competitors in our DeepSeek R1 vs OpenAI o1 analysis.

Can I reproduce the benchmark numbers in the papers using the hosted API?

Not exactly. The V4 paper evaluates the April 2026 preview models, while the API now serves DeepSeek-V4-Pro-0813 (deepseek-v4-pro) and DeepSeek-V4.1-Flash-1017 (deepseek-flash), so scores will differ. The closest setup is reasoning_effort="max" with a max_tokens large enough for the reasoning content; temperature has no effect in thinking mode. Agentic benchmarks depend on tool-use scaffolding that the papers describe but do not ship as a kit. The DeepSeek API getting started tutorial has a minimal client setup.

Sources & Methodology

Last source check: November 12, 2026

Primary sources

Technical references

Context sources

Methodology

Information in this article was checked against the official DeepSeek documentation, the article's cited primary sources, and the deepseekai.guide pricing dataset on the date shown above. Where third-party numbers are quoted, the originating source is linked inline so the reader can verify and re-check.

Data confidence

High for facts cited inline with primary-source links; medium for context narrative.

Editorial note

AI tooling moves quickly. If you spot an outdated number or a broken link, the contact link in the footer is the fastest way to flag it for the editorial team.

Leave a Reply

Your email address will not be published. Required fields are marked *