DeepSeek

Learning Objectives

• Understand why DeepSeek's cost-efficient training methods had such a significant impact on the AI industry
• Distinguish between DeepSeek's current V4.1-Flash (September 2026), the retiring V4-Pro and V4-Flash pair (April 2026), the V3/V3.2 chat models, and the R1 reasoning model
• Know when to use DeepSeek's API or chat interface vs. running the open-weight models locally
• Understand the security and regulatory concerns surrounding DeepSeek

What Is DeepSeek?

DeepSeek is a Chinese AI research lab founded in 2023 by High-Flyer Capital Management, a quantitative hedge fund. In January 2025, DeepSeek released DeepSeek R1 — an open-source reasoning model that matched OpenAI's o1 on major benchmarks — and simultaneously published a research paper claiming the model was trained for approximately $5.6 million, compared to the hundreds of millions spent training comparable US frontier models. As of September 2026 that figure is formally disputed: three US agencies say it excludes the cost of data DeepSeek obtained by distilling US frontier models, and China's foreign ministry calls the accusation groundless. Treat it as a contested vendor claim rather than a settled number.

On September 10, 2026 Anthropic published its own account of the same pattern, naming DeepSeek among seven Chinese labs. Its description of DeepSeek's method is specific: a chain-of-thought extraction pipeline using a cross-session replay attack — saving the reasoning signature Claude returns in place of raw thinking, opening a new session, and eliciting the model to convert that signature back into the full reasoning trace, defeating a control built precisely to prevent this.

Anthropic further alleges that DeepSeek silently relayed its own customers' requests to Claude without telling them, then kept the transcripts as training data. Users who believed they were using a DeepSeek model were receiving Claude's answers, and their sessions — which Anthropic says included sensitive personal and corporate material — reached a third party they had not chosen. Anthropic is a direct competitor and DeepSeek has not publicly responded, so this is an allegation rather than an established fact. It is also a different kind of claim from model theft: it concerns what a lab did with its own paying users' data.

The release triggered immediate global reaction. On January 27, 2025, NVIDIA lost $589 billion in market value in a single day — the largest single-day loss in stock market history. The Nasdaq fell 3.1%, and approximately $1 trillion was wiped from US tech stocks. Markets subsequently recovered fully, with NVIDIA reaching a $5 trillion market cap by October 2025. But the episode forced a fundamental reexamination of assumptions about the capital requirements for frontier AI.

DeepSeek offers three primary products in 2026:

1. DeepSeek V4.1-Flash (September 2026) — the current model: a 552 billion-parameter multimodal Mixture-of-Experts with a 1 million-token context window, MIT-licensed, now serving the traffic of both April models
2. DeepSeek V4-Pro and V4-Flash (April 2026) — the outgoing MoE foundation models, also MIT-licensed with 1 million-token context; V4-Flash is retired and V4-Pro is being wound down
3. DeepSeek Chat (V3/V3.2) — the previous-generation general-purpose conversational models, still widely used
4. DeepSeek R1 — the chain-of-thought reasoning model designed for complex mathematical, coding, and logical tasks

The DeepSeek Model Family

V4.1-Flash — The Small Model That Replaced the Flagship (September 2026)

On September 10, 2026, DeepSeek released V4.1-Flash and used the release notes to retire its own flagship. V4-Flash and V4-Flash-Vision-Exp were withdrawn the same day, their model names left as aliases pointing at the new one. From noon Beijing time on September 14, every request to the V4-Pro endpoint is served by V4.1-Flash and billed at the Flash price, because DeepSeek says its own testing showed the smaller model now beats V4-Pro on performance, cost and speed together.

The architecture is where the saving comes from. V4.1-Flash is a natively multimodal mixture-of-experts model with a 552 billion parameter backbone, arranged as a Causal Encoder-Decoder: a 20-layer causal encoder followed by a 20-layer decoder, where the decoder's global key-value cache is projected from the encoder's final hidden states rather than derived layer by layer. That lets it activate only 8 billion parameters per token while reading input and 16 billion while generating, which is what makes long input-heavy agent work cheap. Combined with FP4 caching and a sparse-attention scheme DeepSeek calls Compressed Sparse Attention 2, the persistent key-value cache falls to 890 bytes per token — roughly a quarter of V4-Flash. It was pretrained on 45 trillion multimodal tokens and carries a reasoning-effort setting the caller can dial from 1 to 100.

The weights are public on Hugging Face, ungated, under a plain MIT license — no revenue threshold, no user-count trigger, no excluded territory. That is worth stating plainly because it is increasingly the exception rather than the rule among Chinese open-weight releases: Kimi K3 and Qwen3.8 Max both ship under custom terms with commercial conditions attached.

Where the trade actually lands. V4.1-Flash is smaller than the model it replaces, so this is not a straightforward capability jump — DeepSeek's argument is that the architecture recovers enough that the size no longer buys you anything at this price. Two things follow. V4-Pro remains the larger model and, at 1.6 trillion parameters, still the largest open-weight model available under fully permissive terms; anyone who wants that specific artifact should download it rather than assume the endpoint will keep serving it. And V4-Pro is text only, so for the first time DeepSeek's default endpoint understands images — the retirement is a capability gain as well as a price cut.

V4 — Frontier-Adjacent Open Weights (April 2026)

On April 24, 2026, DeepSeek released two new MoE foundation models — V4-Pro and V4-Flash — both MIT-licensed and downloadable from Hugging Face. These are the first DeepSeek models to ship with a 1 million-token context window, matching Claude and Gemini's industry-leading context length.

V4-Pro is the flagship: 1.6 trillion total parameters with 49 billion active per token in an MoE architecture. The full model weighs roughly 865 gigabytes on Hugging Face. According to DeepSeek's accompanying paper, V4-Pro trails state-of-the-art frontier models by approximately 3 to 6 months on most benchmarks. It was the largest open-weights model at release; Moonshot's 2.8-trillion-parameter Kimi K3 and Alibaba's 2.4-trillion-parameter Qwen3.8 Max have since passed it on size, though both carry custom licenses, leaving V4-Pro the largest open-weight model available under fully permissive MIT terms. Notable efficiency gain: V4-Pro uses approximately 27% of V3.2's FLOPs and 10% of the KV cache at 1 million-token context — meaningful both for training cost and on-device inference.

V4-Flash is the smaller sibling at 284 billion total / 13 billion active parameters (about 160 gigabytes), targeting the cost-efficient inference tier. Pricing puts V4-Flash below GPT-5.4 Nano, Claude Haiku, and the Gemini Flash variants.

API pricing on platform.deepseek.com, as the V4 pair winds down:

• V4-Pro: 66 cents in and one dollar ninety-eight out per million tokens off-peak, one dollar thirty-two and three dollars ninety-six at peak. The 0813 build reached general availability on August 12, 2026 at a flat 44 cents / 87 cents, ending a preview that had run since April; the peak/off-peak rates above replaced that on August 16. These are the last V4-Pro rates — from September 14, 2026 the endpoint bills at the V4.1-Flash price instead
• V4-Flash: 22 cents in and 66 cents out off-peak, 44 cents and one dollar thirty-two at peak. Retired on September 10, 2026; the name still resolves, but the model behind it is V4.1-Flash

Both models remain downloadable from Hugging Face for self-hosting, and are carried by third-party providers (Together.ai, Fireworks AI, Groq), which is the route to keep using them once DeepSeek's own endpoints stop serving them.

V4-Flash-Vision — The First Multimodal V4 (August 2026)

On August 31, 2026, DeepSeek published DeepSeek-V4-Flash-Vision-Exp, the first multimodal model in the V4 family, as ungated weights under an MIT license. It adds vision modules to the V4-Flash architecture and continues training to unlock visual understanding, while holding roughly level with text-only V4-Flash on text-agent tasks.

The gains are concentrated in multimodal agent work. DeepSeek's own comparison table reports ApexBench rising from 26.2 to 36.5 and Agents' Last Exam from 25.2 to 27.3 against text-only V4-Flash — though the earlier scores come from a model that simply ignores images in the input, so some of that lift measures the absence of a capability rather than the quality of a new one. Against Anthropic's Opus 4.8, DeepSeek reports itself ahead on Agents' Last Exam (27.3 to 25.7) and ZeroBench (35.0 to 34.0), and behind on ApexBench (36.5 to 39.4) and most text-agent benchmarks including NL2Repo (57.7 to 69.7).

Read the license per checkpoint, not per family. This one is plain MIT, and that is genuinely permissive — but it licenses only this checkpoint. Across a single week in late August 2026, Chinese labs shipped weights under terms ranging from plain MIT through Apache 2.0 to custom licenses carrying revenue and user-count triggers, sometimes from the same vendor. The Exp suffix is also load-bearing: this is an experimental release shipping reference inference code rather than a supported product, with a vLLM recipe targeting a four-GPU GB300 node — capable hardware most readers will rent rather than own.

DeepSpec — Open-Source Speculative Decoding (June 2026)

In June 2026, DeepSeek open-sourced DeepSpec, a full-stack codebase for training and evaluating speculative-decoding algorithms, along with three drafting modules — DSpark, DFlash, and Eagle3 — that bolt onto the V4 models to speed up text generation. Speculative decoding lets a small "draft" model propose several tokens at once for the larger model to verify in parallel, cutting the number of expensive full-model passes. DeepSeek's accompanying paper reports generation speedups in the range of 60 to 85% on its V4 checkpoints. The release continues the lab's pattern of publishing its efficiency tooling openly rather than keeping it proprietary — the same posture that made its FP8 training and Multi-Token Prediction work influential across the industry.

First Outside Funding Round — Open-Source AGI Mandate

DeepSeek closed its first outside venture round in June 2026, raising approximately 50 billion yuan (about $7.4 billion) — the first external capital in the lab's history. Commercial investors were led by Tencent, which put in 10 billion yuan, and battery maker CATL at 5 billion yuan, with NetEase and JD.com also participating; founder Liang Wenfeng contributed another 20 billion yuan from his own holdings.

The structure is as notable as the size. Commercial backers accepted five-year lockups and no voting rights. The only investor granted governance rights and a direct stake was Beijing's National Artificial Intelligence Industry Investment Fund, China's state-backed strategic AI vehicle (distinct from the chip-focused "Big Fund") — an arrangement that concentrates control with the founder and the Chinese state rather than diluting it across the cap table. The round places DeepSeek alongside frontier US labs by valuation, even though its training spend remains a fraction of theirs.

The defining feature of the round is the mandate Liang has set publicly with investors: DeepSeek will keep developing open-source models and pursue artificial general intelligence as its core goal, resisting the usual pressure to chase near-term commercialization. That posture is unusual at this scale — most frontier US labs draw their largest checks from corporate cloud partners (OpenAI and Microsoft, Anthropic and Amazon) and treat AGI claims with strategic ambiguity. DeepSeek is doing the opposite, and doing it with one of the largest state-aligned bets on AGI to date outside the United States.

V3.2-Speciale — Competition-Grade Performance

In December 2025, DeepSeek briefly released V3.2-Speciale, a competition-focused model that achieved extraordinary results:

• IMO gold medal with 35 out of 42 points
• 10th place at IOI (International Olympiad in Informatics)
• 96.0% on AIME — exceeding GPT-5-High's 94.6%

The model was available via API only until December 15, 2025, before being discontinued due to extreme computational costs. V3.2-Speciale demonstrated that DeepSeek's training methodology could produce models competitive with or exceeding the very best US frontier models on the hardest reasoning tasks.

Core Capabilities

DeepSeek V3.2 — Efficient Frontier Chat

DeepSeek V3.2 is DeepSeek's general-purpose chat model. Its technical architecture uses Mixture-of-Experts (MoE) with 671 billion total parameters, activating approximately 22 billion for any given query — delivering frontier-adjacent performance at dramatically lower inference cost.

Key capabilities:

• Multi-turn conversation and instruction following
• Code generation, debugging, and explanation across major languages
• Mathematical reasoning and problem-solving
• Document summarization and analysis
• 128K context window for long documents

DeepSeek R1 — Open-Source Reasoning

DeepSeek R1 uses a process similar to OpenAI's chain-of-thought training — the model explicitly "thinks through" problems before producing a final answer. You can see this reasoning process in the response (shown as a collapsible "thinking" section in the chat interface).

The updated R1-0528 variant adds JSON output and function-calling capabilities, making it more practical for agentic and structured-output applications.

This approach excels at:

• Mathematics: Competition-level math problems, proofs, calculations
• Coding: Debugging complex programs, writing algorithms from specifications
• Logic puzzles: Multi-step reasoning chains, formal logic
• Scientific reasoning: Physics, chemistry, biology problem-solving

Hybrid Thinking Mode

DeepSeek's chat interface supports a thinking mode toggle — switching between quick responses (V3.2 chat mode) and extended reasoning (R1 mode). This is similar to Claude's extended thinking or ChatGPT's reasoning mode — useful for hard problems, unnecessary for simple queries.

Pricing & Access

DeepSeek's pricing changed direction twice in a month, and the second turn undid the first. On August 16, 2026 the API moved off flat rates onto peak and off-peak billing — peak hours run from 1 to 4 and from 6 to 10 UTC, with off-peak charged at half the peak rate — and the headline numbers rose sharply: V4-Pro output went from 87 cents per million tokens to one dollar ninety-eight off-peak and three dollars ninety-six at peak, while cached input rose roughly 1,100 percent.

Then on September 10 the flagship changed rather than the rate card. V4.1-Flash lists at 60 cents per million output tokens off-peak against V4-Pro's one dollar ninety-eight, so the price of DeepSeek's default endpoint fell by about 70 percent without any headline price cut — the cheap model simply became the only model. The peak and off-peak structure stayed.

DeepSeek remains far below the US frontier tiers on output pricing. Whether it is once again the outright cheapest option depends on which US budget tier you compare against and when you last checked, so treat that as a live question rather than a settled one — and pick DeepSeek for open weights, self-hosting and capability per dollar rather than on an assumption about the bottom of the market.

The Training Cost Story

DeepSeek's most significant contribution to the field may not be the model itself, but the training methodology paper. DeepSeek published that V3 was trained on approximately 2,000 H800 GPUs (lower-spec than the H100s used for US frontier models due to export restrictions) and cost ~$5.9 million in GPU compute — compared to estimates of $50–100 million or more for comparable US models.

The techniques that enabled this efficiency:

• FP8 training precision — reducing memory requirements without significant quality loss
• Mixture-of-Experts architecture — routing tokens to specialized sub-networks
• Multi-Token Prediction (MTP) — predicting multiple future tokens simultaneously, improving training efficiency
• DualPipe pipeline parallelism — reducing communication overhead in distributed training

These innovations have since influenced training approaches across the industry.

Security Concerns

In early 2025, cloud security firm Wiz discovered a publicly accessible DeepSeek database containing over 1 million sensitive records — including chat histories, API keys, and backend details — with zero authentication. The database was exposed to anyone on the internet without any login required. DeepSeek secured the database after Wiz reported the issue, but the incident raised serious questions about the company's security practices.

Government Bans & Restrictions

DeepSeek's Chinese origin and data handling practices have led to significant restrictions:

• Italy blocked DeepSeek in January 2025, citing GDPR violations
• Banned on government devices in South Korea, Australia, Taiwan, and Texas
• Restricted in US House of Representatives offices, NASA, US Navy, and the Pentagon
• Multiple other governments have issued advisories against using DeepSeek for sensitive work

These restrictions apply to DeepSeek's cloud services — running the open-weight models locally on your own infrastructure is not affected by any ban.

Strengths

• Cost efficiency: Frontier-class performance at a fraction of the training and inference cost of US competitors
• Open-weight MIT license: Freely downloadable, fine-tunable, and commercially usable with no restrictions
• R1 reasoning: First open-source reasoning model to match OpenAI o1 — a genuine breakthrough for open AI
• V3.2-Speciale results: IMO gold and 96% AIME demonstrate competition-grade capability exceeding some US frontier models
• Distilled variants: 1.5 billion–70 billion distilled R1 models run on consumer hardware while retaining strong reasoning
• Competitive API pricing: well below the US frontier tiers, and the September 2026 move to V4.1-Flash cut the default endpoint's output price by roughly 70 percent — 60 cents per million tokens off-peak
• Transparent thinking: R1's reasoning chain is visible, helping users understand how the model arrived at answers

Limitations & Considerations

• Data privacy: API and web usage routes data to Chinese servers — use local models for sensitive data
• Security track record: The Wiz database exposure (1 million+ records with zero authentication) raises concerns about operational security
• Censored content: DeepSeek refuses to discuss politically sensitive topics (Taiwan, Tiananmen Square, etc.) — more restrictive than US models in these areas
• Weaker misuse guardrails than the political censorship suggests: the two are not the same axis, and DeepSeek is tighter on one and looser on the other. In a campaign published by Palo Alto Networks' Unit 42 in July 2026, an operator wired DeepSeek into an open-source agent framework and had it enumerate targets, select vulnerabilities, pull public exploit code, and run the attacks with almost no further human input across roughly 460 targets. Unit 42 reports that OpenAI's provider-side controls refused the same requests and disabled the associated account. If your evaluation assumes a frontier-lab refusal layer sits behind the API, verify that assumption rather than inheriting it
• Government bans: Blocked or restricted on government systems in multiple countries — check your organization's policy before using cloud services
• Infrastructure reliability: DeepSeek's own servers have experienced capacity issues during peak demand — third-party API hosts often more reliable
• Not best-in-class for all tasks: Trailing slightly behind the very latest US frontier models (Claude Opus 4.7, GPT-5.5) on complex instruction-following and nuanced writing tasks

Best Use Cases

When to choose alternatives:

• Privacy-sensitive cloud use without local infrastructure → Mistral Le Chat (EU servers)
• Broadest capability ceiling → Claude Opus 4.7, GPT-5.5
• Source-cited research → Perplexity
• Workplace productivity integration → Microsoft 365 Copilot or Google Workspace AI

Getting Started

1. Go to chat.deepseek.com — free with an email account
2. Toggle Think mode on for a math or coding problem and observe the extended reasoning process
3. For local deployment: install Ollama and run ollama run deepseek-r1:7b (7 billion distilled R1 — runs on most consumer GPUs)
4. For API access: visit platform.deepseek.com — pricing is well below the US frontier tiers, but check the live rate card, since peak and off-peak rates now differ and both changed on August 16, 2026

Key Takeaways

• DeepSeek's R1 release in January 2025 was a watershed moment — the first open-source reasoning model competitive with OpenAI's o1, trained for ~$5.9 million rather than hundreds of millions, triggering a $589 billion single-day NVIDIA stock drop
• V3.2-Speciale (Dec 2025) achieved IMO gold and 96.0% AIME, exceeding GPT-5-High — demonstrating that DeepSeek can match or beat the best US models on competition-grade reasoning
• MIT license means all DeepSeek model weights are freely downloadable, fine-tunable, and deployable on-premise — eliminating data privacy concerns associated with their cloud API
• Security concerns are real: Wiz discovered a publicly accessible database with 1 million+ records and zero authentication; multiple governments have banned DeepSeek on official devices
• DeepSeek V4 shipped April 2026 in two MIT-licensed variants — V4-Pro (1.6 trillion total / 49 billion active, 1 million-token context, the largest open-weight model under a fully permissive license) and V4-Flash (284 billion total / 13 billion active) — both priced below the equivalent US frontier tiers. V4-Pro left preview as the 0813 build on August 12, 2026 with a large price cut, then reversed course on August 16 when DeepSeek moved the whole API to peak/off-peak billing at sharply higher rates
• Both were superseded on September 10, 2026 by V4.1-Flash, a 552 billion-parameter multimodal MoE that activates 8 billion parameters per token while reading input and holds its key-value cache to 890 bytes per token. V4-Flash retired the same day and V4-Pro's endpoint routes to V4.1-Flash from September 14, which cut the default output price by roughly 70 percent and gave the default endpoint vision for the first time. The lesson is that a smaller model displaced a larger one on cost and speed rather than on raw scale — V4-Pro is still the bigger artifact and still downloadable, so keep the weights if you need that specific model
• In June 2026 DeepSeek open-sourced DeepSpec, a speculative-decoding toolkit (drafting modules DSpark, DFlash, and Eagle3) that speeds up V4 text generation — continuing the lab's pattern of publishing its efficiency tooling openly
• DeepSeek closed its first outside funding round in June 2026, raising roughly 50 billion yuan (about $7.4 billion) from Tencent, CATL, NetEase, and JD.com, with founder Liang Wenfeng contributing 20 billion yuan himself; commercial backers took five-year lockups and no voting rights, and only Beijing's state-backed National AI Industry Investment Fund received governance rights — a structure that keeps control with the founder and the Chinese state while the lab stays publicly committed to open-source AGI