DeepSeek
Hangzhou lab known for cost-efficient open-weight models whose January 2025 R1 release shook markets; now past $1B in revenue and preparing a STAR Market IPO.
DeepSeek is the Chinese lab that made efficiency the story of 2025. Spun out of the quantitative hedge fund High-Flyer and founded by Liang Wenfeng in 2023, it released DeepSeek-V3 in December 2024 and the reasoning model R1 in January 2025. R1 matched leading US reasoning models at a fraction of the reported training cost, and the resulting selloff wiped hundreds of billions of dollars off Nvidia’s market value in a single day.
Models
See DeepSeek models. DeepSeek V4 Pro / V4.1 Flash-1017 (Oct 17 2026) is a 552B-parameter sparse MoE under the MIT license with 1M context, plus Search & Fetch APIs at $0.75 per 1,000 requests. DeepSeek open-sourced DeepInfer (Apache 2.0), its serving stack, with kernels for Nvidia SM100 and Huawei Ascend 950, and released the DeepSeek Harness agent runtime.
Business
Annualized revenue crossed $1B in September 2026 as API prices rose. It reportedly raised about $15B (100B yuan) at ~$74B pre-money from 15+ investors including Tencent, CATL, Meituan and Geely, targeting a Shanghai STAR Market listing in early 2027 (not confirmed by DeepSeek). Funds reportedly go to an Inner Mongolia data center with 160,000+ Huawei Ascend chips.
Security and policy
The US Department of Homeland Security ordered DeepSeek technology removed from federal civilian systems. DeepSeek models turn up in attacker tooling: the ARTEX bank breach ran partly on DeepSeek v4.1-flash, CrowdStrike says the VAPOR PANDA campaign ran ARTEX mainly on DeepSeek (no sign of DeepSeek’s involvement), and DeepSeek was found in PHP2Shell attack infrastructure. Unit 42 says the CL-STA-1238 group drove the Cairn framework with DeepSeek V4 Flash and Pro, and the FBI’s Nov 13 2026 ARTEX alert says the DeepSeek Flash family dominates attacker-run ARTEX instances. A Nov 2026 weight analysis by a former OpenAI researcher claims DeepSeek V4 shows signs of training on export-banned Nvidia Blackwell chips (the researcher’s claim, not independently verified). Its Harness had a critical sandbox-escape CVE (see DeepSeek Harness (dsh)).
Timeline
Every dated event in the wiki that involves DeepSeek, newest first.
- Vals finds frontier models pick civilian targets in simulated strikes, especially when told it is a game
The evaluation startup tested models from OpenAI, Anthropic, Google, DeepSeek and others on simulated drone and satellite imagery. Models often flagged civilians and surrendering soldiers as targets, refusals were uncommon, game framing often got past them, and no request was blocked by a safety classifier. Anthropic's models did worst: in a school-targeting case Claude Opus 5.5 picked the school in 69% of ordinary trials and 72% of game-framed ones (Vals findings, via TechCrunch).
The FBI warned that ARTEX, an open-source framework that turns any LLM into an intrusion agent, is being used by criminals and nation-states against government and critical sectors, and keeps spreading through forks, archived copies and a commercial version after its author pulled the repository. It says attacker-run instances mostly use Chinese models, led by the DeepSeek Flash family, and cites Censys's count of over 1,600 exposed consoles.
- Weight analysis claims top Chinese open models were trained on export-banned Nvidia Blackwell chips
Former OpenAI researcher Will DePue (Unemployment Labs) says the released weights of models from DeepSeek, Moonshot AI, MiniMax and Ant Group carry fingerprints of Blackwell training hardware, including Kimi K3 and DeepSeek V4. TechCrunch notes this would not show the models were trained only on Nvidia chips or that export law was broken, since offshore data centers are a common route. The claims are the researcher's, not independently verified; the story runs alongside the Megaspeed smuggling inquiry.
Palo Alto Networks Unit 42 says Chinese-speaking attackers it tracks as CL-STA-1238 ran 25 operations in Aug-Sep 2026 with a customized fork of the open-source Cairn pentest framework, driven by Zhipu GLM 5.2 and DeepSeek V4 Flash/Pro. A Chinese-language preamble told the models the work was authorized and forbade refusal; neither model refused. Unit 42 says Cairn confirmed access or data retrieval at 10 organizations.
CrowdStrike says China-nexus VAPOR PANDA ran the ARTEX agent framework, mainly on DeepSeek, against 200+ organizations in 14 countries in 11 days, confirming post-exploitation at nine.
An attacker rooted 25 servers, three in a U.S. government-affiliated environment; Claude Code, DeepSeek and Kimi were found in the attack infrastructure.
- DeepSeek: Second private round reported closing (July-October)
Second private round reported closing (July-October); investors reported as 15+ incl. Tencent, CATL, Meituan, Geely, Midea, NetEase, GigaDevice; Shanghai STAR Market IPO targeted early 2027 with CITIC Securities as adviser. Neither report quotes DeepSeek; Reuters says details could still change.
Ordered DeepSeek technology removed from federal civilian IT systems and blocked future purchases of its products and services, including third-party products built on DeepSeek technology (exceptions possible).
Sources: TechCrunch, “US orders federal agencies to remove China's DeepSeek AI from their systems”; WSJ (via The Neuron)
- DeepSeek open-sources DeepInfer
Apache-2.0 serving stack for V4.1-Flash with kernels for Nvidia SM100 and Huawei Ascend 950.
Sources: GitHub
- DeepSeek releases V4.1 Flash-1017
Minor revision of the 552B V4.1 Flash (released Sep 10) under MIT with 1M context, same API name and rate card; Search & Fetch tools reported as separately priced.
Full article: DeepSeek modelsSources: The Next Web, “DeepSeek launches V4.1-Flash and retires V4-Pro, its flagship model” (for the base V4.1 Flash release); Hugging Face
- DeepSeek: second round nearing, with Tencent, CATL, GigaDevice, SMIC and Geely named
Reported (not confirmed): second round nearing, with Tencent, CATL, GigaDevice, SMIC and Geely named; funds an Inner Mongolia data center with 160,000+ Huawei Ascend chips
Sources: Seoul Economic Daily (citing Caijing)
CrowdStrike tied a breach of Shinhan Bank and Yegaram Savings Bank to an operator running the ARTEX agent on DeepSeek, GLM and Grok models.
- DeepSeek: Pre-IPO round with Tencent and CATL (reported
Pre-IPO round with Tencent and CATL (reported; Hong Kong IPO targeted early 2027)
Sources: Bloomberg (via AI Weekly)
- DeepSeek previews DeepSeek Harness
MIT-licensed Node.js agent runtime where everything is a plugin; later hit by a critical sandbox-escape CVE.
Sources: The New Stack
Connections
DeepSeek and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →
What links here
AI lab Alibaba (Qwen) · Meituan (LongCat)
Product ARTEX · DeepSeek Harness (dsh) · Tencent Yuanbao Search API
Incident ARTEX breach of Korean banks · CL-STA-1238 Cairn campaign · PHP2Shell intrusions · VAPOR PANDA ARTEX campaign
Model family DeepSeek models · GPT-OSS
Government body FBI · US Department of Homeland Security
Concept Inference · Open-weight vs. closed models
Person Liang Wenfeng
Publication The Wall Street Journal