DeepSeek

Company profile

• Founded: 2023 (incubated by the quantitative hedge fund High-Flyer 幻方量化)
• Headquarters: Hangzhou
• Main models:
  • DeepSeek-V3 (general-purpose), DeepSeek-R1 (reasoning), plus V3/R1 derivatives
  • MoE architecture + reinforcement-learning training path are DeepSeek’s distinctive technical signatures
  • Open-source weights (MIT License) + open API
• Business model: API (at extremely low prices) + open-source ecosystem-driven
• Positioning: one of the most globally influential Chinese open-source LLM providers

Policy document snapshot

Type	Document	Link	Subpage
Usage policy	DeepSeek service agreement / privacy policy	chat.deepseek.com (footer)	usage-policy
Model cards	DeepSeek-V3 / R1 technical reports	GitHub + arXiv	model-card
Safety framework	Limited public documentation; TC260-003 compliance as baseline	limited public documentation	safety-framework
Transparency report	CAC filings (filed / published)	cac.gov.cn	transparency-report
Red-team disclosures	Partial disclosure in papers and system cards	github.com/deepseek-ai	red-team-disclosures

Regulatory-compliance posture

• China:
  • Public-facing API / chat services subject to CAC algorithm filing
  • Whether publishing open weights itself requires filing: regulatory practice is ambiguous (service / weights dual path)
  • Subject to TC260-003 baseline
• United States:
  • The early-2025 “DeepSeek moment” triggered technical shock and policy backlash
  • BIS export-control updates; CFIUS review
  • Bans on government devices at the state and federal-agency level
• European Union:
  • Open-source models used to provide downstream services in the EU → GPAI obligations apply
  • Italy’s Garante issued a temporary ban and investigation in January 2025
• South Korea / Taiwan: restrictions on use on government devices

Deep dive: maximalist open source + minimal self-regulation

DeepSeek’s distinctive combination

DeepSeek is the global AI company closest to “no standalone self-regulation narrative”:

• No RSP / Preparedness / FSF-equivalent documents
• No explicit safety-principles statement (unlike Google AI Principles)
• Extremely thin Usage Policy (a few hundred words, essentially stating “do not use for unlawful purposes”)
• Safety-evaluation disclosure: only partial sections of technical reports
• All governance commitments = CAC filing materials (not publicly disclosed)

Compared with Mistral’s open source:

• Mistral: open source + EU compliance + GPAI CoP signatory
• DeepSeek: open source + Chinese compliance + no international industry self-regulation commitment
• Both substitute open source for transparency, but Mistral adds an EU institutional layer; DeepSeek does not

Global impact of the “DeepSeek moment” (2025-01)

The January 2025 release of DeepSeek R1 was seismic:

• Performance comparable to OpenAI o1 but with training cost estimated at an order of magnitude lower
• Open source → anyone can download, fine-tune, and deploy
• Triggered U.S. equity-market volatility (AI stocks sold off) and White House attention
• Accelerated the Trump administration’s deregulation + compute-competition narrative in EO 14179 + the AI Action Plan

Three-fold pressure on global AI governance:

1. Questioning of compute thresholds: if $5M can train an o1-class model, how should the EU 10^25 FLOP and California SB 53 10^26 FLOP thresholds be defined? (The DeepSeek-V3 technical report is arxiv 2412.19437; R1 is arxiv 2501.12948.)
2. Open source is irreversible: once DeepSeek weights are released, any “pause” or “withdrawal” becomes meaningless
3. Self-regulation cannot bind open-source companies: all self-regulation frameworks assume the company controls model access — open source breaks the assumption

International regulatory responses

• 2025-01 Italy Garante: the first EU DPA temporary ban on DeepSeek (GDPR inadequacy)
• 2025-02 South Korea, Taiwan, Australia: government-device bans
• 2025-03 U.S. federal agencies: DoD, Commerce, FCC, etc. ban DeepSeek
• 2025-04 Texas, Virginia, and other states: state government-device bans
• 2025-10 EU AI Office: DeepSeek included in GPAI systemic-risk review (operations outside the EU but with model weights flowing in)

CAC filing’s distinctive practical challenge

DeepSeek’s compliance architecture:

• chat.deepseek.com public-facing service → CAC algorithm filing
• API service → likewise filed
• Open weights (HuggingFace / GitHub) → regulatory practice is ambiguous:
  • CAC has not clarified whether “publishing weights” constitutes “providing a service”
  • Q2 2025 internal CAC discussions considered whether open models must be filed
  • As of 2026-04, no definitive rule

This is an open boundary of Chinese AI regulation, and DeepSeek is the single most important test case.

Compared with other Chinese companies

Dimension	DeepSeek	Alibaba Qwen	Baidu ERNIE	ByteDance Doubao
Open-source strategy	Maximalist open source (MIT)	Primarily open (Apache 2.0)	Mixed (4.5 open / 5.0 closed)	Partial open
Self-regulation documents	Least	Few	Medium (official blog)	Few
Government relations	Independent (quantitative hedge fund)	National team (Alibaba Cloud national platform)	National team (Apollo)	Commercial + geopolitical
International reach	Greatest (“DeepSeek moment”)	Large (Qwen global downloads)	Medium	Medium (TikTok + Doubao)
Regulatory pressure	Highest (multi-country bans)	Medium	Medium	High (TikTok + Doubao)

Company posture, in brief

• Technical-hacker aesthetic + low-price open-source operating strategy
• Extremely few public AI-governance position papers; the company behaves as if “the model itself is the best statement”
• Open source’s practical pressure on global AI governance: weight releases are irreversible
• Quantitative-fund background gives DeepSeek greater political independence than Alibaba / Baidu (national team) or ByteDance (commercial + geopolitical)