新增: DeepSeek 全部 33 篇官方论文(原文 PDF 走 Git LFS + 中文精读笔记与索引)

大模型/deepseek/_scripts/guide.txt

@@ -0,0 +1,152 @@
+DeepSeek Research Papers: V4 to V2 Reading Guide Skip to content No results Home Models Tutorials Comparisons Guides Reviews API Tools About DeepSeek AI Guide H
+ome Models Tutorials Comparisons Guides Reviews API Tools About DeepSeek AI Guide Search Menu 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. April 24, 2026 News Quick facts Quick fa
+cts Topic A Guide to DeepSeek Research Papers From V2 to V4 Format News brief Models discussed V4-Pro, V4-Flash, V4, V3.2 Reading time 11 min read (~2,519 words
+) Last reviewed: April 30, 2026 · Source · View sources On this page 12 sections Why DeepSeek’s papers matter more than most The DeepSeek paper timeline at a gl
+ance The foundation papers (2024) DeepSeek LLM — the scaling-law study DeepSeekMoE and DeepSeek-V2 — the architecture base DeepSeek-Coder and DeepSeekMath DeepS
+eek V3 — the efficiency thesis in paper form DeepSeek R1 — reasoning without human demonstrations DeepSeek V3.2 — the sparse attention bridge DeepSeek V4 — the 
+current paper Architecture contributions Positioning against closed frontier models How the papers connect: a suggested reading order Using the papers as a deve
+loper, not a reader What the papers do not tell you Frequently asked questions Where can I find the official DeepSeek research papers? What does the DeepSeek V4
+ paper actually introduce? How much did it cost to train DeepSeek V3 and R1? Is the DeepSeek R1 paper really a reasoning trace breakthrough? Can I reproduce the
+ benchmark numbers in the papers using the hosted API? Sources & Methodology Primary sources Technical references Context sources Methodology Data confidence Ed
+itorial note A Guide to DeepSeek Research Papers From V2 to V4 News · April 24, 2026 · By DS Guide Editorial If you are trying to understand why DeepSeek V4 use
+s only 10% of the KV cache of V3.2, 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 and Hugging Face, from the original DeepSeek LLM scaling-law work in Janua
+ry 2024 to the DeepSeek-V4 technical report released on April 24, 2026. For each paper we note the model it describes, the single architectural idea it contribu
+ted, 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 Most frontier AI labs ship products and a blog post. DeepSeek ships products, open weights, and a detai
+led technical report — usually on arXiv, usually the same day, usually with enough implementation detail to reproduce the architecture. That is how the lab beca
+me a reference point in 2025: DeepSeek drew wide attention in 2025 when its R1 release delivered near industry-leading reasoning performance – for allegedly a f
+raction of the price. The V4 preview, released on April 24, 2026, continues that pattern. The paper presents two Mixture-of-Experts language models — DeepSeek-V
+4-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. 
+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 prim
+ary sources. The DeepSeek paper timeline at a glance Every paper below is first-party DeepSeek-AI work. I have limited the list to the releases that matter for 
+understanding the current V4 generation. For a wider bibliography, the community-maintained DeepSeek Papers collection on Hugging Face is the most complete sing
+le 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) 2024-02-0
+5 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 241
+2.19437 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 2026-04-24 Hugging Face PDF Hybrid CSA + HCA attention, 1M context by default The foundation papers (2
+024) 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 7
+B 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 wa
+y 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 . D
+eepSeekMoE and DeepSeek-V2 — the architecture base DeepSeekMoE (January 2024) introduced the fine-grained expert split that every later DeepSeek MoE model inher
+its. 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 eco
+nomical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of train
+ing 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 De
+epSeekMath Two specialist papers from early 2024 matter for the V4 story. DeepSeek-Coder introduced the code-pretraining recipe that fed later general-purpose r
+eleases. DeepSeekMath introduced Group Relative Policy Optimisation (GRPO) , the RL algorithm that R1 would later use to learn reasoning without human demonstra
+tions. 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 V
+3 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 6
+71B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attenti
+on (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. Furthermore, DeepSeek-V3 pioneers an auxiliary-loss-free strategy for lo
+ad balancing and sets a multi-token prediction training objective for stronger performance. DeepSeek-V3 was pre-trained on 14.8 trillion diverse and high-qualit
+y tokens, followed by Supervised Fine-Tuning and Reinforcement Learning stages. The paper’s most-quoted number is the training cost: Combined with 119K GPU hour
+s for 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 t
+he H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M. That is the number that Reuters and everyone else converted to “about $5.6M.” T
+he paper is careful to note this excludes prior research and ablations — a caveat many press summaries drop. Our DeepSeek V3 review covers how the claims held u
+p in production. DeepSeek R1 — reasoning without human demonstrations R1 is the paper that turned DeepSeek from “efficient lab” to “strategic concern” in Wester
+n 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 fa
+cilitates the emergent development of advanced reasoning patterns, such as self-reflection, verification, and dynamic strategy adaptation. The trained model ach
+ieves 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. The $5.576M figure in the V3 paper is a separa
+te 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 tra
+ined on top of DeepSeek-V3-Base and DeepSeek-R1 reuses non-reasoning data from DeepSeek-V3 SFT data. The two papers are a pair. Our DeepSeek R1 review and the D
+eepSeek R1 vs OpenAI o1 comparison pair well with the paper. DeepSeek V3.2 — the sparse attention bridge V3.2 is the under-rated paper. Released in December 202
+5, it is the one that lays the groundwork for V4’s efficiency claims. The paper introduces DeepSeek Sparse Attention (DSA), an efficient attention mechanism tha
+t substantially reduces computational complexity while preserving model performance in long-context scenarios. By implementing a robust reinforcement learning p
+rotocol 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 Collegiate Programming Contest. V3.2’s DSA is instantiated on top of MLA, not as a replacement for it. The V3.2 paper is the most detailed ac
+count 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 cur
+rent paper The V4 technical report, titled DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence , is hosted as a PDF in the V4-Pro repositor
+y on Hugging Face. It is the primary source for every V4 claim. Architecture contributions The paper’s headline contribution is a new hybrid attention stack. De
+epSeek-V4 designs a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-c
+ontext efficiency. In the 1M-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 incorporates Manifold-Constrained Hyper-Connections (mHC) to strengthen conventional residual connections, enhancing stability of signal p
+ropagation across layers while preserving model expressivity. Two more items are worth extracting from the report: Training precision. FP4 + FP8 Mixed: MoE expe
+rt parameters use FP4 precision; most other parameters use FP8. FP4 experts are unusual and are part of why V4 trained at the cost it did. Three reasoning-effor
+t modes. DeepSeek-V4-Pro and DeepSeek-V4-Flash both support three reasoning effort modes. These map to the API’s non-thinking default, reasoning_effort="high" w
+ith a thinking flag, and reasoning_effort="max" . In the API this returns reasoning_content alongside the final content . Positioning against closed frontier mo
+dels The report is unusually candid about where V4 sits. Simon Willison extracted the key quote: “Through the expansion of reasoning tokens, DeepSeek-V4-Pro at 
+maximum reasoning effort demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performanc
+e falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails current-generation frontier models by approximately 3 
+to 6 months.” That is a self-description the press did not have to spin. On specific benchmarks the paper reports, V4-Pro (reasoning_effort=max) wins on LiveCod
+eBench (93.5 — a new open-model high), Codeforces rating (3206 — DeepSeek reports this places the model roughly 23rd among human contest participants), Apex Sho
+rtlist (90.2), SimpleQA-Verified at 57.9 against open and near-open models, and a proof-perfect 120/120 on Putnam-2025 formal reasoning under the hybrid informa
+l-formal pipeline. Verify any individual number against the PDF before you quote it — frontier benchmark standings shift monthly. How the papers connect: a sugg
+ested reading order If you have a weekend and want the architecture thread, read in this sequence: DeepSeek-V2 for MLA and DeepSeekMoE — the foundation. DeepSee
+k-V3 for auxiliary-loss-free load balancing, multi-token prediction, and FP8 training. DeepSeek-R1 for the RL-only reasoning recipe (which uses GRPO from DeepSe
+ekMath). DeepSeek-V3.2 for DSA as an attention-efficiency prototype. DeepSeek-V4 for the CSA + HCA hybrid and the full mHC / Muon / FP4 training stack. If you o
+nly have an evening, read the V4 report and V3 report back to back. Together they cover roughly 80% of the current generation’s ideas. The DeepSeek benchmarks 2
+026 page has the cross-model numbers. Using the papers as a developer, not a reader Every V3 paper onward describes a model you can call through POST /chat/comp
+letions , the OpenAI-compatible endpoint at https://api.deepseek.com . The current V4 model IDs are deepseek-v4-pro and deepseek-v4-flash . Legacy IDs ( deepsee
+k-chat , deepseek-reasoner ) remain accepted until 2026-07-24 15:59 UTC, when they retire; both route to deepseek-v4-flash in the meantime. A minimal Python cal
+l 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 DSA section."
+}], reasoning_effort="high", extra_body={"thinking": {"type": "enabled"}}, ) print(resp.choices[0].message.reasoning_content) print(resp.choices[0].message.cont
+ent) 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. 
+The web chat at chat.deepseek.com keeps session history; the API does not. Context window is 1,000,000 tokens by default on both V4 tiers, with output up to 384
+,000 tokens. Temperature guidance from DeepSeek’s docs: 0.0 for code and maths (matching the paper’s evaluation setup), 1.0 for data tasks, 1.3 for general chat
+ and translation, 1.5 for creative writing. JSON mode is designed to return valid JSON, not guaranteed. Prompt explicitly with the word “json” plus a small sche
+ma example, and set max_tokens high enough to avoid truncation. For costs, V4-Flash runs at $0.0028 (cache hit) / $0.14 (cache miss) / $0.28 (output) per 1M tok
+ens; V4-Pro at promotional $0.435 / $0.87 (75% off through 2026-05-31; list $0.0145 / $1.74 / $3.48). A 1M-call workload with a 2,000-token cached system prompt
+, a 200-token user message and a 300-token response on V4-Flash works out to $5.60 cached input + $28 uncached input + $84 output = $117.60 total . On V4-Pro th
+e same workload is $29 + $348 + $1,044 = $1,421 . Check the DeepSeek API pricing page before committing — Preview-window rates can move. For the full API surfac
+e, start with the DeepSeek API documentation and the DeepSeek API getting started tutorial. What the papers do not tell you 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-referenc
+e against independent benchmark aggregators before making a procurement decision. Training-cost figures exclude prior research. The V3 paper says this explicitl
+y. The $5.576M number is the official run, not the full R&D bill. Licensing is per-model. V4-Pro, V4-Flash, V3.2, V3.1 and R1 publish weights under MIT. Some ol
+der releases — V3 base, Coder-V2, VL2 — split code (MIT) from weights (a separate DeepSeek Model License). Check the specific Hugging Face repo if licensing mat
+ters to your deployment. For broader independent context, the DeepSeek latest updates feed tracks post-publication corrections and community replications. DeepS
+eek AI Guide is an independent resource and is not affiliated with DeepSeek or its parent company. Model IDs, pricing and API behaviour change; check the offici
+al DeepSeek documentation and pricing page before committing to a production decision. Frequently asked questions Where can I find the official DeepSeek researc
+h papers? DeepSeek publishes technical reports on arXiv and as PDFs in the matching Hugging Face repository. The V4 report lives inside the V4-Pro model repo; V
+3 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 introduc
+es a hybrid attention stack — Compressed Sparse Attention plus Heavily Compressed Attention — that the authors claim cuts single-token inference FLOPs to 27% an
+d 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 in th
+e PDF linked from the 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 
+$2/hour, totalling about $5.576M — explicitly excluding prior research and ablations. DeepSeek’s separately disclosed R1 training cost was around $294,000 (Reut
+ers, September 2025). The two figures belong to different models and should not be combined. Our DeepSeek V3 review unpacks the cost claim. Is the DeepSeek R1 p
+aper 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 mo
+del. 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 host
+ed API? Partially. You can match most reasoning and coding benchmarks by setting model="deepseek-v4-pro" , reasoning_effort="max" , temperature=0.0 , and a max_
+tokens large enough for the reasoning content. 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: April 30, 2026 Primary sources Official DeepSeek API document
+ation Endpoints, request/response shape, model IDs Last checked: April 30, 2026 Pricing DeepSeek API pricing page Per-1M-token rates, billing units, cache-hit p
+ricing Last checked: April 30, 2026 Model card DeepSeek on Hugging Face Weights, license, usage notes per model Last checked: April 30, 2026 Repository Hugging 
+Face: DeepSeek Papers community collection Community-maintained bibliography of DeepSeek papers Last checked: April 30, 2026 Technical references Technical repo
+rt DeepSeek-V3 technical report (arXiv:2412.19437) V3 architecture and training methodology Last checked: April 30, 2026 Technical report DeepSeek-R1 paper (arX
+iv:2501.12948) R1 reasoning approach, distillation results Last checked: April 30, 2026 Technical report DeepSeek LLM scaling-law paper Scaling-law study for 7B
+ and 67B dense models Last checked: April 30, 2026 Technical report DeepSeekMoE paper Fine-grained expert specialisation architecture Last checked: April 30, 20
+26 Technical report DeepSeek-Coder paper Code-specialised base model recipe Last checked: April 30, 2026 Technical report DeepSeekMath / GRPO paper Group Relati
+ve Policy Optimisation introduction Last checked: April 30, 2026 Technical report DeepSeek-V2 technical report MLA + MoE architecture foundation for later model
+s Last checked: April 30, 2026 Technical report DeepSeek-V3 technical report V3 efficiency thesis, training-cost figure Last checked: April 30, 2026 Technical r
+eport DeepSeek-R1 technical report RL-only reasoning recipe and benchmark results Last checked: April 30, 2026 Technical report DeepSeek-V3.2 technical report (
+DSA) Sparse attention bridge before V4 Last checked: April 30, 2026 Technical report DeepSeek V4 technical report CSA + HCA hybrid attention, FLOP / KV cache cl
+aims Last checked: April 30, 2026 Context sources News Reuters: V3 ~$5.6M training cost (Jan 2025) Conversion of GPU-hour figure to dollar headline Last checked
+: April 30, 2026 News Reuters: R1 $294,000 training cost (Sep 2025) Distinct R1 cost figure separate from V3 Last checked: April 30, 2026 Analysis Simon Williso
+n: V4 launch-day writeup Extracted key quotes from V4 technical report Last checked: April 30, 2026 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 nu
+mbers 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 i
+s the fastest way to flag it for the editorial team. DG DS Guide Editorial Editorial team DS Guide Editorial is the in-house editorial team at deepseekai.guide.
+ We write practical, source-cited reviews and tutorials for the DeepSeek model family — V4-Pro, V4-Flash, R1, V3, Coder — and the broader frontier-AI ecosystem.
+ Every article carries a Last reviewed date and a Sources section so you can verify what we said and when. Last reviewed: April 30, 2026 · Editorial methodology
+: view sources Related articles View all in News → News DeepSeek Latest Updates — What Shipped With V4 Preview (April 2026) April 24, 2026 News DeepSeek V4 Rele
+ase Date: Pro and Flash Land as an Open-Weight Preview April 24, 2026 News DeepSeek Company News: V4, Funding Talks and What Shifted in 2026 April 24, 2026 News
+ DeepSeek Benchmarks 2026: What the V4 Numbers Actually Say April 24, 2026 News DeepSeek Funding News: Inside the $300M First Round April 24, 2026 News DeepSeek
+ Partnerships in the V4 Era: Who’s Actually Shipping April 24, 2026 Leave a Reply Cancel Reply Your email address will not be published. Required fields are mar
+ked * Name * Email * Website Add Comment * Save my name, email and website in this browser for the next time I comment. Post Comment DeepSeek AI Guide Independe
+nt reviews, tutorials, and comparisons for the DeepSeek AI ecosystem — no affiliation, no hype. Written and maintained by working AI practitioners. Explore Mode
+ls Tutorials Comparisons Guides Reviews Alternatives Resources API Docs Free Tools Alternatives News Company About Contact Editorial Guidelines Sitemap Independ
+ence notice: DeepSeek AI Guide is an independent resource and is not affiliated with, endorsed by, or sponsored by DeepSeek or its parent company. All trademark
+s and product names are the property of their respective owners and are used for identification purposes only (nominative fair use). © 2026 DeepSeek AI Guide. A
+ll rights reserved. Privacy Policy Terms of Use Disclaimer Cookie Policy Editorial Guidelines