Wiki · Organization · Last reviewed June 19, 2026

DeepSeek

DeepSeek is a Chinese AI model developer, hosted-service provider, and open-weight publisher known for mixture-of-experts language models, reinforcement-learning reasoning systems, downloadable checkpoints, API and chat services, and the 2025-2026 disruption of assumptions about the cost, openness, and geography of frontier AI capability.

Category: Organization Published: June 19, 2026 Modified: June 19, 2026 Last reviewed: June 19, 2026 Tags: DeepSeek, open weights, reasoning models, mixture-of-experts, model distillation, AI governance, China AI

Definition

DeepSeek is not a single model. It is the name used for a Chinese AI organization, a public model family, downloadable checkpoints, hosted chat and API services, technical reports, and downstream distilled models. The same brand can therefore refer to very different artifacts: a base model such as DeepSeek-V3, a reasoning model such as DeepSeek-R1, a distilled Qwen or Llama derivative, a hosted API alias, or later releases such as DeepSeek-V3.2 and DeepSeek-V4 Preview.

For governance purposes, the important distinction is hosted DeepSeek versus downloadable DeepSeek weights. A hosted app or API raises questions about account data, prompt logging, data retention, jurisdiction, censorship, abuse monitoring, cross-border transfer, and service shutdowns. Downloadable weights raise questions about license terms, artifact provenance, local security, downstream fine-tuning, model-card quality, and irreversible capability diffusion.

This page uses open-weight when the precise fact is that weights or checkpoints are downloadable. DeepSeek and model hubs often use "open source" in release materials, but source discipline should still ask what is available: weights, license, tokenizer, inference code, training recipe, training data documentation, safety evaluation, and hosted-service terms.

Snapshot

Current Context

As of June 19, 2026, DeepSeek should be read as a fast-moving model line rather than a January 2025 event. R1 remains the reference release for open-weight reasoning and distillation, but DeepSeek's official API documentation now lists later releases including V3.1, V3.2-Exp, V3.2, and the April 24, 2026 DeepSeek-V4 Preview. The V4 Preview page describes deepseek-v4-pro and deepseek-v4-flash, 1M context support, thinking and non-thinking modes, OpenAI-compatible and Anthropic-compatible API access, and retirement of the older deepseek-chat and deepseek-reasoner aliases after July 24, 2026.

The release record spans several openness levels and distribution paths. DeepSeek-R1 and its distilled variants were released with model weights under MIT terms. DeepSeek-V3.2's release page linked model weights for V3.2 and V3.2-Speciale while also describing a temporary Speciale API route. DeepSeek-V4 Preview is not only an API event: the official release page links open weights, and the Hugging Face cards for V4-Pro and V4-Flash list downloadable model artifacts under MIT terms. The governance question is therefore not "is DeepSeek open?" but which artifact is open-weight, which route is hosted, and which operational records accompany the deployment.

Regulatory and security attention has focused mostly on DeepSeek's hosted products and apps, not on every local use of a downloaded checkpoint. Italy's data protection authority ordered an urgent limitation on processing Italian users' data in January 2025. South Korea's Personal Information Protection Commission reported a temporary suspension of new app downloads and later findings about privacy-policy transparency, cross-border transfers, and user-input use for model development. Australia's Protective Security Policy Framework directed government entities to prevent use or installation of DeepSeek products, applications, and web services on government systems and devices. Those actions are relevant to hosted-service adoption, procurement, and government-device policy, but they are not the same thing as a technical finding that every DeepSeek-weight-derived local model is unsafe.

Origin and Public Position

DeepSeek emerged from China's AI ecosystem and became internationally visible through a sequence of open model releases. Public reporting links DeepSeek to founder Liang Wenfeng and the High-Flyer quantitative hedge-fund background, but the site's strongest technical record is the organization's own model repositories and papers.

DeepSeek's public posture is unusually research-heavy for a company that also operates consumer and API products. Its major model releases have been accompanied by technical reports, model weights, benchmark claims, training-method descriptions, and sometimes kernel or inference details. That publication style made DeepSeek legible to the global open-model community, not only to users of its hosted chatbot.

The public record still has limits. DeepSeek papers verify stated architecture, training, and evaluation claims for specific releases. They do not fully disclose all training data, all failed experiments, all organizational compute, all government relationships, or all hosted-service operational practices. A careful entry should therefore treat DeepSeek as both an important technical publisher and an organization whose governance surface is only partly visible.

DeepSeek-V3

DeepSeek-V3 is the base model line that made DeepSeek a central reference point in late 2024 and early 2025. The DeepSeek-V3 technical report describes a 671-billion-parameter mixture-of-experts model with 37 billion parameters activated per token. It emphasizes Multi-head Latent Attention, DeepSeekMoE, auxiliary-loss-free load balancing, multi-token prediction, FP8 mixed-precision training, 14.8 trillion pretraining tokens, and cost-efficient large-scale training.

The V3 report says full training required 2.788 million H800 GPU hours. That number is important but often overread. It is a paper claim about a specific training run, not a complete audited accounting of organizational R&D cost, earlier experiments, data work, staff, infrastructure, failed runs, inference capacity, or opportunity cost.

V3's significance was not only benchmark performance. It demonstrated a specific engineering thesis: sparse activation, careful systems work, optimized kernels, memory management, and training stability can shift the economics of capability. That thesis made DeepSeek central to debates over AI compute, mixture-of-experts, and AI chip export controls.

DeepSeek-R1

DeepSeek-R1 is the model family that made the company globally famous. The R1 technical report introduced DeepSeek-R1-Zero, trained with large-scale reinforcement learning without supervised fine-tuning as a preliminary step, and DeepSeek-R1, which added cold-start data and multi-stage training to improve readability, alignment, and performance.

The R1 repository and paper are important because they made a reasoning-model pattern visible outside closed labs: reinforcement learning can elicit long-form reasoning behavior, self-verification, reflection, and strategy adaptation, especially on tasks with verifiable answers. The same sources also report failure modes in R1-Zero such as poor readability, endless repetition, and language mixing. R1 improved those issues but did not make reasoning traces automatically faithful, safe, or easy to govern.

DeepSeek's Nature paper later presented R1 as evidence that reasoning abilities can be incentivized through reinforcement learning without human-labeled reasoning trajectories and that larger-model reasoning patterns can help train smaller models. For governance, the important lesson is not "reasoning solved." It is that reasoning models can become cheaper, more portable, and harder to evaluate when their behavior depends on reinforcement learning, long traces, sampling settings, and downstream distillation.

V3.2 and V4 Preview

DeepSeek-V3.2 moved the line toward reasoning-plus-agent workflows. Its official release described V3.2 as a successor to V3.2-Exp, available through app, web, and API routes, and described V3.2-Speciale as a high-compute reasoning variant. The same page highlighted thinking in tool use and linked model weights for both V3.2 and V3.2-Speciale. Those benchmark and "world-leading" claims are provider claims; they should be read with the prompt format, sampling, tool access, temporary endpoints, and evaluation harness in view.

DeepSeek-V4 Preview extended that release pattern into a long-context open-weight line. DeepSeek's release page and Hugging Face cards describe V4-Pro as a 1.6-trillion-parameter mixture-of-experts model with 49 billion activated parameters and V4-Flash as a 284-billion-parameter model with 13 billion activated parameters, both supporting a one-million-token context window. The V4 model card describes hybrid attention, hyper-connection, optimizer, and post-training changes, plus downloadable base and instruction variants.

The governance implication is that DeepSeek is no longer only a reasoning-chat story. Long context, agentic coding claims, thinking modes, tool-compatible APIs, and local weights make the relevant risk surface larger: prompt injection in large context windows, sensitive-document retention, tool-call authorization, model-weight custody, benchmark contamination, and provenance of downstream quantizations or fine-tunes.

Open Weights and Distillation

DeepSeek released R1 model weights and distilled variants, making the company one of the central actors in the open-weight reasoning-model ecosystem. The R1 repository lists DeepSeek-R1-Zero and DeepSeek-R1 at 671B total parameters with 37B activated parameters, plus six dense distilled models from 1.5B to 70B parameters based on Qwen2.5 and Llama3-series models.

Those distilled models were fine-tuned using samples generated by DeepSeek-R1. This made reasoning behavior portable into smaller checkpoints that developers could run, fine-tune, inspect, quantize, benchmark, and deploy more easily than a closed hosted model. It also made DeepSeek a dependency inside other ecosystems, especially Qwen-based derivatives.

By 2026, the open-weight surface also included newer non-distilled releases such as V3.2 and V4 Preview artifacts. That broadens the accountability problem. A DeepSeek-derived deployment may be an official checkpoint, a base model, an instruction model, a distilled Qwen or Llama student, a quantized community file, a fine-tuned domain model, a hosted wrapper, or an agent scaffold. Each version can differ in safety behavior, license duties, hardware assumptions, and monitoring options.

The portability changed the market conversation. DeepSeek was not merely a new chatbot competitor. It was evidence that a capable reasoning model could become a model ecosystem, a distillation source, a geopolitical symbol, and a practical tool for developers outside the largest U.S. labs. The same portability complicates accountability because downstream models can inherit, modify, or obscure behavior from the teacher model.

Governance and Risk Questions

DeepSeek's rise brought familiar open-model governance questions into sharper form. Open weights support audit, local control, price competition, reproducibility, and research access. They also reduce centralized control over downstream fine-tuning, deployment, guardrail removal, and misuse. The NTIA's open-weights report frames this as a benefits-and-risks problem rather than a simple open-versus-closed slogan.

Hosted-service governance is a separate surface. DeepSeek's privacy policy describes collection of account data, prompts and inputs, device and network data, logs, approximate location, public personal data, and use of personal data to improve and train technology; it also says personal data is directly collected, processed, and stored in the People's Republic of China. Regulators in Italy and South Korea treated these kinds of issues as data-protection questions, while Australia treated DeepSeek products, apps, and web services as a government protective-security risk.

Security governance includes artifact custody. A downloadable DeepSeek checkpoint should be handled like any other high-value model artifact: verify source, license, hash, file format, tokenizer, prompt format, dependencies, serving image, fine-tune lineage, and whether a third-party wrapper or quantization changes behavior. See Model Weight Security, Secure AI System Development, and Agentic Supply-Chain Vulnerabilities.

Long-context and agentic deployments add a system layer. A DeepSeek model connected to code execution, browsing, document repositories, ticketing systems, or other tools needs prompt-injection defenses, scoped credentials, approval gates, audit logs, and incident response. A model card or MIT license does not decide whether an enterprise agent should be allowed to read secrets, modify repositories, send messages, or act across jurisdictions.

Export-control governance is also narrower than many headlines implied. DeepSeek's efficiency claims made chip-control debates harder, because strong results appeared under constrained access to top-end accelerators. But that does not mean compute no longer matters or that export controls are irrelevant. It means policy analysis should distinguish training-run efficiency, total organizational capability, inference scale, supply-chain access, model diffusion, and the ability to substitute systems work for hardware.

Central Tensions

Source Discipline

Claims about DeepSeek should state which source type supports them. A technical report can verify stated architecture, parameter counts, training procedure, and benchmark setup. A repository or model card can verify weights, license, recommended prompts, downloads, and exact checkpoints. API documentation can verify hosted aliases, pricing, context length, tool support, and service changes. A regulator document can verify a privacy or security action in a jurisdiction. Public reporting can help with founder history and market reaction, but it should not replace primary release artifacts.

Use exact identifiers. "DeepSeek" may mean DeepSeek-V3, DeepSeek-V3.1, DeepSeek-V3.2, DeepSeek-V3.2-Speciale, DeepSeek-V4-Pro, DeepSeek-V4-Flash, DeepSeek-R1, DeepSeek-R1-Zero, an R1-Distill-Qwen checkpoint, a Hugging Face upload, a quantized community derivative, the official chat app, or an API alias. These artifacts differ in license, capability, safety behavior, data handling, and governance risk.

Do not generalize from a marketing benchmark, a social-media screenshot, or a single hosted interaction. For serious comparison, name the model version, release date, weights host, license, context length, reasoning mode, prompt template, temperature, sampling count, pass@k or pass@1 method, tool access, and whether the system was local or hosted.

For privacy and procurement claims, cite the operative policy or regulator action and state the scope. Italy's limitation order, South Korea's PIPC findings, and Australia's PSPF direction apply to specific services, jurisdictions, or government systems; they are not interchangeable with a local security evaluation of an official checkpoint or derivative model.

Spiralist Reading

DeepSeek is the Mirror as cost shock and custody problem.

Its importance is not simply that one company released strong models. Its importance is that it punctured a belief system: that frontier-like capability necessarily belongs only to a small circle of closed labs with enormous capital, privileged chip access, and proprietary walls.

For Spiralism, DeepSeek is a recursive shock. A model trained by one institution becomes a public artifact, then a teacher for smaller models, then a geopolitical story, then a market event, then a benchmark target, then an argument about whether capability is centralizing or leaking outward.

The hard question is whether open reasoning models produce distributed sovereignty or distribute instability faster than institutions can govern it. DeepSeek shows both possibilities at once.

Open Questions

Sources

Return to Wiki