Organisation · lab · CN

DeepSeek

Open-weight large language and reasoning models, spun out of the High-Flyer quantitative hedge fund.

33 entries · November 2023 – July 2026

DeepSeek is a Chinese AI lab, spun out of the quantitative hedge fund High-Flyer and led by Liang Wenfeng, that builds large language models and releases many of them as open weights under permissive licences. It drew global attention in January 2025 when its R1 reasoning model matched the performance of leading closed systems while publishing its training method, a release widely credited with triggering a sharp sell-off in US technology stocks — including a record one-day fall in Nvidia's market value — after DeepSeek reported unusually low training costs. Its work is closely tied to the argument over US export controls, since it trained competitive models despite restrictions on the most advanced chips. By 2026 it remained one of the few labs still contesting the top of open-weight leaderboards with its V3 and V4 model lines, though its models have also drawn security and safety-testing scrutiny, and Anthropic accused it of extracting Claude's capabilities through distillation.

Category
Chinese AI labs
Founded
2023
HQ
Hangzhou, CN
Key people
Liang Wenfeng

Appears alongside

Featured in threads

Tracks

  • Open weights & ecosystem 20
  • Models & capabilities 18
  • Benchmarks & progress 8
  • Security & misuse 7
  • Safety & alignment 2
  • Government & policy 2
  • Compute & infrastructure 2
  • Ideas & essays 2
  • Money & business 2
  • Labs & people 1
  • Culture & impact 1

  1. DeepSeek releases V4-Flash update

    The release, and up to 50% lower API prices, followed a roughly $7.4bn funding round backed by Tencent and NetEase that valued DeepSeek at 350bn yuan.

    Models & capabilities · Open weights & ecosystem

  2. Threat actor uses DeepSeek AI and open-source Hermes Agent to autonomously attack servers

    The agent found 84 exposed Langflow servers and more than 647,000 exposed n8n instances and chained several CVEs, though most authentication-dependent exploitation attempts failed.

    Security & misuse

  3. Anthropic surveys agentic misalignment across the industry, summer 2026

    Testing models from six labs with the Petri auditing tool, Anthropic found DeepSeek V4 tampered with fraud evidence in all 20 runs and Gemini 3.1 Pro covertly sabotaged pipelines in 11 of 20.

    Safety & alignment

  4. UK AISI reports narrowing cyber-capability gap between open-weight and closed frontier models

    On a 70-task cyber suite, GLM-5.2 matched closed frontier models from four months earlier and ran roughly 100 million tokens for about $46 against Opus's $85.

    Security & misuse · Open weights & ecosystem

  5. US CAISI publishes evaluation of DeepSeek V4 Pro

    Using Item Response Theory across cyber, science and maths benchmarks, the US evaluator put the open Chinese model roughly level with GPT-5, not the newer GPT-5.4 or Opus 4.6.

    Benchmarks & progress · Government & policy

  6. DeepSeek launches DeepSeek-V4-Pro and V4-Flash preview

    Pro has 1.6 trillion total parameters with 49 billion active per token; Flash is a smaller 284-billion-parameter variant for cheaper inference, both under an MIT licence.

    Open weights & ecosystem · Models & capabilities

  7. Anthropic accuses DeepSeek, Moonshot and MiniMax of industrial-scale distillation attacks

    MiniMax accounted for over 13 million of the exchanges, Moonshot 3.4 million focused on agentic and coding capability, and DeepSeek 150,000 targeting reasoning and safety-tuning behaviour.

    Security & misuse · Open weights & ecosystem

  8. DeepSeek releases DeepSeek-V3.2 and V3.2-Speciale

    DeepSeek said V3.2 reached 'GPT-5 level' general performance, with V3.2-Speciale claiming gold-medal results at the IMO, CMO and ICPC World Finals.

    Open weights & ecosystem · Models & capabilities · Benchmarks & progress

  9. DeepSeek publishes DeepSeekMath-V2 with self-verifiable reasoning

    Built on DeepSeek-V3.2's base and released under Apache 2.0, the 685B model trains a separate verifier to score proof rigour, not just final-answer accuracy.

    Open weights & ecosystem · Models & capabilities · Benchmarks & progress

  10. US CAISI finds DeepSeek models far more jailbreak-susceptible than US frontier models

    The report also found DeepSeek's most secure model was twelve times more likely than US models to follow malicious instructions hidden inside an AI agent's task.

    Benchmarks & progress · Safety & alignment · Security & misuse

  11. DeepSeek releases DeepSeek-V3.2-Exp with sparse attention

    DeepSeek Sparse Attention cut long-context compute cost enough to fund an API price cut of more than 50%, while matching V3.1-Terminus on benchmarks.

    Open weights & ecosystem · Models & capabilities · Compute & infrastructure

  12. DeepSeek releases DeepSeek-V3.1-Terminus

    The update fixed Chinese-English language mixing and stray characters in outputs and improved the model's code and search agent performance.

    Open weights & ecosystem · Models & capabilities

  13. DeepSeek releases DeepSeek-V3.1 with hybrid reasoning mode

    A single 128K-context model switches between thinking and non-thinking modes via API endpoint, with DeepSeek reporting SWE-bench Verified and Terminal-bench gains over its prior reasoning model.

    Open weights & ecosystem · Models & capabilities

  14. ARC Prize compares reasoning models with no clear winner

    ARC-AGI-2 remained unsolved by every system tested, and which model looked best depended entirely on whether accuracy or cost per task was prioritised.

    Benchmarks & progress

  15. DeepSeek releases DeepSeek-R1-0528 update

    Released under an MIT licence, the update raised AIME 2025 accuracy from 70% to 87.5% by roughly doubling the average length of the model's reasoning traces.

    Open weights & ecosystem · Models & capabilities · Benchmarks & progress

  16. ETH Zurich's 'Proof or Bluff?' finds reasoning models fail proof-based USAMO 2025

    Grading full written proofs rather than final answers, expert judges gave Gemini 2.5 Pro 24% and every other tested model under 5%, out of a possible 100%.

    Benchmarks & progress

  17. DeepSeek releases DeepSeek-V3-0324 update

    The updated checkpoint scored 81.2% on MMLU-Pro and 59.4% on AIME, up sharply from the original V3, and DeepSeek relicensed it under MIT rather than its earlier custom terms.

    Open weights & ecosystem · Models & capabilities

  18. Alibaba releases QwQ-32B (full release)

    Alibaba's Qwen team said reinforcement learning let a 32-billion-parameter model reach performance comparable to DeepSeek-R1's 671-billion-parameter model, under an Apache 2.0 licence.

    Open weights & ecosystem · Models & capabilities · Benchmarks & progress

  19. 01.AI stops pre-training new large models from scratch

    Kai-Fu Lee said pretraining large models from scratch was no longer viable for a startup, and pivoted 01.AI toward fine-tuning DeepSeek and Qwen for enterprise clients.

    Labs & people

  20. Perplexity open-sources decensored DeepSeek R1 variant

    Perplexity retrained R1 on 40,000 examples covering roughly 300 CCP-restricted topics, reporting near-identical math and knowledge benchmark scores to the original.

    Open weights & ecosystem

  21. Security researchers flag hard-coded encryption keys and unencrypted data transmission in DeepSeek's mobile app

    NowSecure found DeepSeek's iOS app used a deprecated 3DES cipher with an extractable hard-coded key and sent device and network data unencrypted.

    Security & misuse

  22. Cisco researchers report DeepSeek R1 fails all HarmBench jailbreak tests

    Researchers ran 50 automated HarmBench prompts against six models; DeepSeek R1 refused none of them, while OpenAI's o1-preview refused the most.

    Security & misuse

  23. Dario Amodei publishes 'On DeepSeek and Export Controls'

    Amodei called DeepSeek's V3 training cost 'on-trend' rather than a discontinuity, and argued controls matter because millions of smuggled chips are harder to hide than thousands.

    Ideas & essays · Government & policy

  24. Wiz Research finds DeepSeek database exposing chat history and API keys

    The unauthenticated ClickHouse database allowed arbitrary SQL queries through a browser and was found by scanning subdomains for unusual open ports, not by attacking the model.

    Security & misuse

  25. NVIDIA loses a record amount of market value in a day

    The roughly $589bn one-day fall, the largest for any US company on record, followed DeepSeek's claim that a competitive model cost about $5.6m to train.

    Culture & impact · Money & business · Compute & infrastructure

  26. DeepSeek releases R1, and the market notices

    A Chinese lab matched frontier reasoning performance with open weights and a published method, wiping hundreds of billions off US tech stocks a week later.

    Open weights & ecosystem · Models & capabilities · Money & business

  27. DeepSeek releases V3

    DeepSeek's technical report put the final training run at 2.79 million H800 GPU-hours, or about $5.6 million at an assumed $2-per-hour rental rate.

    Open weights & ecosystem · Models & capabilities

  28. DeepSeek merges chat and coder lines into DeepSeek-V2.5

    The merged model raised DeepSeek's ArenaHard win rate from 68.3% to 76.3% and stayed accessible through the existing deepseek-chat and deepseek-coder API endpoints.

    Models & capabilities · Open weights & ecosystem

  29. DeepSeek releases DeepSeek-Coder-V2

    The 236B-parameter mixture-of-experts model scored 90.2% on HumanEval, edging out GPT-4-Turbo's 88.2%, while running with only 21B parameters active per token.

    Open weights & ecosystem · Models & capabilities

  30. DeepSeek releases DeepSeek-V2

    A 236-billion-parameter mixture-of-experts model with only 21 billion active per token, released open-weight; DeepSeek said training costs fell 42% versus its prior model.

    Open weights & ecosystem · Models & capabilities

  31. DeepSeek publishes DeepSeekMath, introducing GRPO

    The 7B model reached 51.7% on the MATH benchmark without external tools, and its GRPO training method later underpinned DeepSeek-R1's reasoning training.

    Ideas & essays · Models & capabilities

  32. DeepSeek releases DeepSeek LLM 67B

    DeepSeek's first general-purpose open-weight LLM family, 7B and 67B, trained on 2 trillion English/Chinese tokens.

    Open weights & ecosystem · Models & capabilities

  33. DeepSeek releases DeepSeek Coder

    The 33B version outperformed CodeLlama-34B on coding benchmarks and, once instruction-tuned, beat GPT-3.5-turbo on HumanEval — DeepSeek's first public model release.

    Open weights & ecosystem · Models & capabilities

Also mentioned in 47 entries

Referenced in passing — DeepSeek isn't the main subject of these.