# deepseek: side quests

> the pursuits around the flagship — the philanthropy, the compute-donation program, the edge model lines, open source week, and the raise.

- 2019– — an ordinary little pig (liang, continuing): the whale's quiet philanthropy: rmb 221.38m across 15 charities in 2022 alone, and liang personally giving rmb 138m under the pseudonym 一只平凡的小猪 — 'an ordinary little pig' — disclosed only through high-flyer's own history page. (sources: [official timeline 2008–2022](https://www.high-flyer.cn/history/))
- 2021– — the idle-time compute program (company, continuing): the idle-time compute program: ~27% of fire-flyer ii's 2022 capacity — over 15m gpu-hours — donated to 50+ university and institute labs on a time-sharing scheduler. the same scheduler that time-shares trading and research workloads gives academia the surplus. (sources: [official timeline 2008–2022](https://www.high-flyer.cn/history/), [the hfai blog: the 10k-a100 cluster](https://www.high-flyer.cn/blog/hfai/), [fire-flyer ii stats (8tb/s read, feb-2022 stats)](https://www.high-flyer.cn/en/))
- 2024-02– — the algorithm that escaped (researchers, handed-off): grpo — group relative policy optimization — ships inside deepseekmath in february 2024, becomes the rl engine of r1, and then escapes the lab entirely: kimi k2 thinking and the rest of the open-weights frontier adopted it. the algorithm is now load-bearing infrastructure for the whole field. (sources: [the deepseek-math repo — grpo](https://github.com/deepseek-ai/deepseek-math/), [k2 thinking beats gpt-5 on hle/browsecomp](https://www.linkedin.com/posts/conorgrennan_breaking-kimi-k2-may-be-chinas-second-deepseek-activity-7392592898703278080-E8tl), [second deepseek moment — kimi k2](https://xpert.digital/en/sputnik-the-moment-of-artificial-intelligence/))
- 2024-04– — [prover](https://github.com/deepseek-ai/DeepSeek-Prover-V2) (researchers, continuing): the theorem-proving line: prover-v1.5, then prover-v2 in april 2025 — a 7b worker plus the 671b teacher for lean 4 proofs. (sources: [prover-v2 (7b/671b), lean 4](https://github.com/DeepSeek-AI/DeepSeek-Prover-V2))
- 2024-12– — vl2 and janus (researchers, continuing): the vision-language and multimodal line: vl2 in december 2024, then janus — and janus-pro-7b shipped on january 28, 2025, in the middle of the crash week. the line keeps shipping while the flagship consumes the attention. (sources: [17 jul 2023 founding](https://en.wikipedia.org/wiki/DeepSeek), [v3 dec 2024 → r1 jan 20 → janus jan 28 timeline](https://www.sydney.edu.au/news-opinion/news/2025/01/29/deepseek-ai-china-us-tech.html))
- 2025-02-24 – 2025-03-01 — [open source week](https://github.com/deepseek-ai/open-infra-index) (company, ended): open source week: six daily drops — flashmla, deepep, deepgemm, dualpipe, eplb with profile data, then 3fs and smallpond — the fund's internal infrastructure stack released as public goods. closed the week with the theoretical-545%-margin disclosure. (sources: [open source week index](https://github.com/deepseek-ai/open-infra-index), [the flashmla repo](https://github.com/deepseek-ai/FlashMLA), [545% theoretical](https://www.scmp.com/tech/big-tech/article/3300734/deepseek-shows-power-v3-r1-models-theoretical-545-profit-margin), [english mirror of the disclosure](https://gigazine.net/gsc_news/en/20250303-deepseek-ai-cost-profit/))
- 2025-01– — the consumer app (company, continuing): the consumer app that briefly beat chatgpt — #1 free app in the us and 51 other countries on january 26 — then became the geopolitical lightning rod: banned from government devices across a dozen countries, yet still ~143m mau in china by august 2025, second to doubao's 157m per questmobile. (sources: [official app announcement](https://www.deepseek.com/en/news/deepseek-app/), [app released jan 10 on ios+android](https://mobiledevmemo.com/deepseek-and-consumer-ais-first-mover-disadvantage/), [#1 us app store jan 26](https://techcrunch.com/2025/01/27/deepseek-displaces-chatgpt-as-the-app-stores-top-app/), [italy/taiwan/navy/texas ban roundup](https://techcrunch.com/2025/02/03/deepseek-the-countries-and-agencies-that-have-banned-the-ai-companys-tech/), [questmobile aug 2025: doubao 157m vs deepseek 143m mau](https://www.scmp.com/tech/big-tech/article/3325858/bytedance-chatbot-doubao-still-chinas-most-popular-ai-app-rival-deepseek-loses-users))
- 2025-04– — the ascend hedge (company, continuing): the domestic-chip stack: after the failed ascend training run (ft, august 2025), the posture settles into nvidia for training and ascend for inference — then v3.1's ue8m0-fp8 'for next-generation domestic chips,' v3.2-exp's day-1 ascend/cann support, and the months of ascend optimization that reportedly delayed v4. (sources: [huawei onsite team](https://www.theregister.com/software/2025/08/14/dodgy_huawei_chips_nearly_sunk_deepseeks_next_gen_r2_mode/753616), [ft: ascend training failure delayed r2](https://www.reuters.com/world/china/deepseeks-launch-new-ai-model-delayed-by-huawei-chip-issues-ft-reports-2025-08-14/), [v3.1's ue8m0 fp8 for domestic chips](https://www.theregister.com/2025/08/22/deepseek_v31_chinese_chip_hints/), [v3.2-exp ships day-1 ascend/cann support](https://www.tomshardware.com/tech-industry/deepseek-new-model-supports-huawei-cann), [v4 delayed for ascend-tuned stack (yuyuantantian)](https://www.bloomberg.com/news/articles/2026-04-26/deepseek-v4-delay-shows-shift-to-china-chips-cctv-account-says))
- 2025-05 – 2026 — r2 (company, ended): the model that never shipped: r2 was reported stalled in june 2025 — liang unsatisfied with performance — then reportedly failed on huawei ascend in august. it was never officially announced and never released; the v-line absorbed the flagship slot. (sources: [deepseek-r2 launch stalled; ceo balks](https://www.reuters.com/world/china/deepseek-r2-launch-stalled-ceo-balks-progress-information-reports-2025-06-26/), [ft: ascend training failure delayed r2](https://www.reuters.com/world/china/deepseeks-launch-new-ai-model-delayed-by-huawei-chip-issues-ft-reports-2025-08-14/), [ft: r2 delayed](https://the-decoder.com/deepseek-reportedly-delayed-its-latest-ai-model-after-technical-issues-with-huaweis-ascend-chips/), [r2 status: unannounced, unreleased as of jul-2026](https://www.layer3labs.io/guides/deepseek-r2-explained))
- 2025-11– — [math-v2](https://arxiv.org/html/2511.22570) (researchers, continuing): self-verifiable theorem proving: math-v2 posts gold-level imo-2025 and cmo-2024 results and 118/120 on putnam 2024 — scaled test-time compute pointed at formal math. (sources: [math-v2: imo-25 & cmo-24 gold-level](https://arxiv.org/html/2511.22570), [authors: shao, luo, lu, ren, hu, ye, gou, ma, zhang](https://github.com/deepseek-ai/DeepSeek-Math-V2))
- 2026-05– — the ipo track (company, active): the listing track: alongside the first external round, reporting puts deepseek on star-market ipo preparation — the lab that 'had no financing plans' heading for public markets inside three years. (sources: [the $74b raise and star ipo prep](https://www.hindustantimes.com/technology/ai-startup-deepseek-poised-to-reach-74-billion-valuation-101787815737042.html), [may-8 round launch](https://www.coinlive.com/news/deepseek-s-liang-wenfeng-we-don-t-want-alibaba-s-money-we-ll-contribute))
- 2024-10– — the fund keeps earning (company, continuing): the whale keeps trading: the market-neutral exit in october 2024, then +56.55% mean product return in 2025 — #2 among >rmb-10b chinese quant privates on >rmb 70b aum. the patron is still winning while the lab spends. (sources: [high-flyer 2025: +56.55%, >rmb 70b](https://www.21jingji.com/article/20260114/herald/c28662157ae4f0d9f2c7c944101a35ad.html), [top china quant winds down strategy](https://www.bloomberg.com/news/articles/2024-10-21/top-china-quant-winds-down-strategy-pummeled-by-market-rally), [neutral exit details](https://m.nbd.com.cn/articles/2024-10-19/3596588.html))

AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.

Canonical page: https://hraness.com/deepseek/side-quests


Published by @ben in profile revision 1.
Original document: https://hraness.com/deepseek/side-quests
Original credited author: Hraness
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.

## Structured source and record data

````json
{"anchors":[{"id":"founders","originalUrl":"https://hraness.com/deepseek/side-quests#founders","title":"the founders' quests"},{"id":"lab","originalUrl":"https://hraness.com/deepseek/side-quests#lab","title":"the lab's quests"},{"id":"bench","originalUrl":"https://hraness.com/deepseek/side-quests#bench","title":"the bench's quests"}],"provenance":{"draftingDisclosure":"AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.","originalAuthor":"Hraness","originalRevision":"1eeab468fd4b2eab481f64dae9ad05d643aa46045544eb1786cbafadd0e22cac","originalUrl":"https://hraness.com/deepseek/side-quests","source":"hraness"},"records":[{"date":{"precision":"approximate","text":"2019–"},"id":"record-303f7bfd4b6566fa2879","kind":"deepseek-quests","label":"an ordinary little pig","original":{"dates":"2019–","detail":"the whale's quiet philanthropy: rmb 221.38m across 15 charities in 2022 alone, and liang personally giving rmb 138m under the pseudonym 一只平凡的小猪 — 'an ordinary little pig' — disclosed only through high-flyer's own history page.","id":"qst-charity","owner":"liang","sourceIds":["src-hp-highflyer-history"],"status":"continuing","title":"an ordinary little pig","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-00a84e4dd575df716128"]},{"date":{"precision":"approximate","text":"2021–"},"id":"record-6b960a0870a3505cfa83","kind":"deepseek-quests","label":"the idle-time compute program","original":{"dates":"2021–","detail":"the idle-time compute program: ~27% of fire-flyer ii's 2022 capacity — over 15m gpu-hours — donated to 50+ university and institute labs on a time-sharing scheduler. the same scheduler that time-shares trading and research workloads gives academia the surplus.","id":"qst-idle-compute","owner":"company","sourceIds":["src-hp-highflyer-history","src-hp-hfai-blog","src-hp-highflyer-home"],"status":"continuing","title":"the idle-time compute program","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-00a84e4dd575df716128","source-904e7f9caba419702302","source-afb542d09f0a4ec6ed22"]},{"date":{"precision":"approximate","text":"2024-02–"},"id":"record-458dfb9e62d1fc33eeb4","kind":"deepseek-quests","label":"the algorithm that escaped","original":{"dates":"2024-02–","detail":"grpo — group relative policy optimization — ships inside deepseekmath in february 2024, becomes the rl engine of r1, and then escapes the lab entirely: kimi k2 thinking and the rest of the open-weights frontier adopted it. the algorithm is now load-bearing infrastructure for the whole field.","id":"qst-grpo","owner":"researchers","sourceIds":["src-hp-deepseekmath-github","src-ci-linkedin-kimi","src-ci-xpert-kimi"],"status":"handed-off","title":"the algorithm that escaped","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-298530916a213e82b46c","source-fed440e6b470b245f2e4","source-e8230989327d1388040d"]},{"date":{"precision":"approximate","text":"2024-04–"},"id":"record-4f6d69ccd404d44eeff5","kind":"deepseek-quests","label":"prover","original":{"dates":"2024-04–","detail":"the theorem-proving line: prover-v1.5, then prover-v2 in april 2025 — a 7b worker plus the 671b teacher for lean 4 proofs.","id":"qst-prover","owner":"researchers","sourceIds":["src-hp-prover-github"],"status":"continuing","title":"prover","url":"https://github.com/deepseek-ai/DeepSeek-Prover-V2"},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-ba66c76fcb42fa01110c"]},{"date":{"precision":"approximate","text":"2024-12–"},"id":"record-55c5d6e2353db7195a5e","kind":"deepseek-quests","label":"vl2 and janus","original":{"dates":"2024-12–","detail":"the vision-language and multimodal line: vl2 in december 2024, then janus — and janus-pro-7b shipped on january 28, 2025, in the middle of the crash week. the line keeps shipping while the flagship consumes the attention.","id":"qst-janus","owner":"researchers","sourceIds":["src-hp-wiki-deepseek","src-ci-sydney-explainer"],"status":"continuing","title":"vl2 and janus","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-446a0de49e666c3afad0","source-324be9fc074dbf2ca8bd"]},{"date":{"precision":"approximate","text":"2025-02-24 – 2025-03-01"},"id":"record-777bd9351117e0e9c2bd","kind":"deepseek-quests","label":"open source week","original":{"dates":"2025-02-24 – 2025-03-01","detail":"open source week: six daily drops — flashmla, deepep, deepgemm, dualpipe, eplb with profile data, then 3fs and smallpond — the fund's internal infrastructure stack released as public goods. closed the week with the theoretical-545%-margin disclosure.","id":"qst-openinfra","owner":"company","sourceIds":["src-hp-openinfra","src-hp-flashmla","src-ci-scmp-545-margin","src-ci-gigazine-545"],"status":"ended","title":"open source week","url":"https://github.com/deepseek-ai/open-infra-index"},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-7a9748e200d02457d23f","source-33b6ba38ceadb66b8b98","source-bca184285a64029f43af","source-77c65b38d1a24a2086fa"]},{"date":{"precision":"approximate","text":"2025-01–"},"id":"record-6e08f55aadf319e7b7a7","kind":"deepseek-quests","label":"the consumer app","original":{"dates":"2025-01–","detail":"the consumer app that briefly beat chatgpt — #1 free app in the us and 51 other countries on january 26 — then became the geopolitical lightning rod: banned from government devices across a dozen countries, yet still ~143m mau in china by august 2025, second to doubao's 157m per questmobile.","id":"qst-app","owner":"company","sourceIds":["src-hp-app-news","src-hp-mobiledevmemo","src-hp-techcrunch-appstore","src-hp-techcrunch-bans","src-ci-scmp-doubao-users"],"status":"continuing","title":"the consumer app","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-f99d11cc91e37d3d6dc1","source-da084182120e71f10ea5","source-c5488a0485bbd430b7b9","source-4e9f90a2f630aa66eaf8","source-c1b95e75a54f58c4b212"]},{"date":{"precision":"approximate","text":"2025-04–"},"id":"record-ff8e73f13b188481153d","kind":"deepseek-quests","label":"the ascend hedge","original":{"dates":"2025-04–","detail":"the domestic-chip stack: after the failed ascend training run (ft, august 2025), the posture settles into nvidia for training and ascend for inference — then v3.1's ue8m0-fp8 'for next-generation domestic chips,' v3.2-exp's day-1 ascend/cann support, and the months of ascend optimization that reportedly delayed v4.","id":"qst-ascend","owner":"company","sourceIds":["src-hp-register-huawei","src-hp-reuters-r2ft","src-ci-theregister-v31","src-ci-tomshardware-v32-cann","src-hp-bloomberg-v4delay"],"status":"continuing","title":"the ascend hedge","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-83afeaedd663fa1fd697","source-2ebfdd7106ca4b48e28d","source-badaf978ae8d380ba07a","source-e3ad300520736dd60743","source-a17c609e873002bfa73c"]},{"date":{"precision":"approximate","text":"2025-05 – 2026"},"id":"record-778cb2eb3e4370963a20","kind":"deepseek-quests","label":"r2","original":{"dates":"2025-05 – 2026","detail":"the model that never shipped: r2 was reported stalled in june 2025 — liang unsatisfied with performance — then reportedly failed on huawei ascend in august. it was never officially announced and never released; the v-line absorbed the flagship slot.","id":"qst-r2","owner":"company","sourceIds":["src-hp-reuters-r2june","src-hp-reuters-r2ft","src-ci-decoder-ascend","src-hp-layer3-r2"],"status":"ended","title":"r2","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-11d00c40b93f5574acad","source-2ebfdd7106ca4b48e28d","source-54ed06394dff64e31ee9","source-3c4d06b451215cd65fb0"]},{"date":{"precision":"approximate","text":"2025-11–"},"id":"record-10a7672071257cdfb86c","kind":"deepseek-quests","label":"math-v2","original":{"dates":"2025-11–","detail":"self-verifiable theorem proving: math-v2 posts gold-level imo-2025 and cmo-2024 results and 118/120 on putnam 2024 — scaled test-time compute pointed at formal math.","id":"qst-mathv2","owner":"researchers","sourceIds":["src-hp-mathv2","src-hp-mathv2-github"],"status":"continuing","title":"math-v2","url":"https://arxiv.org/html/2511.22570"},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-83745b9a03c1c5e17116","source-41500b9a68e5088f3bfa"]},{"date":{"precision":"approximate","text":"2026-05–"},"id":"record-c0cc384ab5f4722140f3","kind":"deepseek-quests","label":"the ipo track","original":{"dates":"2026-05–","detail":"the listing track: alongside the first external round, reporting puts deepseek on star-market ipo preparation — the lab that 'had no financing plans' heading for public markets inside three years.","id":"qst-ipo","owner":"company","sourceIds":["src-ci-ht-valuation","src-hp-coinlive-funding"],"status":"active","title":"the ipo track","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-4b725ef1a727495f648b","source-27c1e329da17603af0fd"]},{"date":{"precision":"approximate","text":"2024-10–"},"id":"record-102ec9937e33214fdbef","kind":"deepseek-quests","label":"the fund keeps earning","original":{"dates":"2024-10–","detail":"the whale keeps trading: the market-neutral exit in october 2024, then +56.55% mean product return in 2025 — #2 among >rmb-10b chinese quant privates on >rmb 70b aum. the patron is still winning while the lab spends.","id":"qst-the-whale","owner":"company","sourceIds":["src-hp-21jingji-hf2025","src-hp-bloomberg-neutral","src-hp-nbd-neutral"],"status":"continuing","title":"the fund keeps earning","url":null},"originalUrl":"https://hraness.com/deepseek/side-quests","sourceIds":["source-def3a541e976061d5949","source-7c345f5754ef4659dd7a","source-84d7b54e8e39bb80541e"]}],"sources":[{"author":null,"id":"source-54ed06394dff64e31ee9","note":"ft: r2 delayed; ascend training failed; revert to nvidia","originalId":"src-ci-decoder-ascend","published":{"precision":"month","text":"2025-08"},"publisher":"the-decoder.com","title":"ft: r2 delayed","type":"reporting","url":"https://the-decoder.com/deepseek-reportedly-delayed-its-latest-ai-model-after-technical-issues-with-huaweis-ascend-chips/"},{"author":null,"id":"source-77c65b38d1a24a2086fa","note":"english mirror of the disclosure","originalId":"src-ci-gigazine-545","published":{"precision":"day","text":"2025-03-03"},"publisher":"gigazine.net","title":"english mirror of the disclosure","type":"reporting","url":"https://gigazine.net/gsc_news/en/20250303-deepseek-ai-cost-profit/"},{"author":null,"id":"source-4b725ef1a727495f648b","note":"$74b raise; $7.4b; catl/tencent investors; star ipo prep","originalId":"src-ci-ht-valuation","published":{"precision":"month","text":"2026-09"},"publisher":"hindustantimes.com","title":"the $74b raise and star ipo prep","type":"reporting","url":"https://www.hindustantimes.com/technology/ai-startup-deepseek-poised-to-reach-74-billion-valuation-101787815737042.html"},{"author":null,"id":"source-fed440e6b470b245f2e4","note":"k2 thinking beats gpt-5 on hle/browsecomp; huang \"nanoseconds behind\"","originalId":"src-ci-linkedin-kimi","published":{"precision":"day","text":"2025-11-07"},"publisher":"linkedin","title":"k2 thinking beats gpt-5 on hle/browsecomp","type":"reporting","url":"https://www.linkedin.com/posts/conorgrennan_breaking-kimi-k2-may-be-chinas-second-deepseek-activity-7392592898703278080-E8tl"},{"author":null,"id":"source-bca184285a64029f43af","note":"545% theoretical; 226.75 nodes×8×h800; $87k→$562k","originalId":"src-ci-scmp-545-margin","published":{"precision":"day","text":"2025-03-01"},"publisher":"scmp.com","title":"545% theoretical","type":"reporting","url":"https://www.scmp.com/tech/big-tech/article/3300734/deepseek-shows-power-v3-r1-models-theoretical-545-profit-margin"},{"author":null,"id":"source-c1b95e75a54f58c4b212","note":"questmobile aug 2025: doubao 157m vs deepseek 143m mau","originalId":"src-ci-scmp-doubao-users","published":{"precision":"month","text":"2025-09"},"publisher":"scmp.com","title":"questmobile aug 2025: doubao 157m vs deepseek 143m mau","type":"reporting","url":"https://www.scmp.com/tech/big-tech/article/3325858/bytedance-chatbot-doubao-still-chinas-most-popular-ai-app-rival-deepseek-loses-users"},{"author":null,"id":"source-324be9fc074dbf2ca8bd","note":"v3 dec 2024 → r1 jan 20 → janus jan 28 timeline","originalId":"src-ci-sydney-explainer","published":{"precision":"day","text":"2025-01-29"},"publisher":"sydney.edu.au","title":"v3 dec 2024 → r1 jan 20 → janus jan 28 timeline","type":"reference","url":"https://www.sydney.edu.au/news-opinion/news/2025/01/29/deepseek-ai-china-us-tech.html"},{"author":null,"id":"source-badaf978ae8d380ba07a","note":"v3.1 ue8m0 fp8 for next-gen domestic chips","originalId":"src-ci-theregister-v31","published":{"precision":"day","text":"2025-08-22"},"publisher":"theregister.com","title":"v3.1's ue8m0 fp8 for domestic chips","type":"reporting","url":"https://www.theregister.com/2025/08/22/deepseek_v31_chinese_chip_hints/"},{"author":null,"id":"source-e3ad300520736dd60743","note":"v3.2-exp day-1 ascend/cann/vllm support","originalId":"src-ci-tomshardware-v32-cann","published":{"precision":"day","text":"2025-09-29"},"publisher":"tomshardware.com","title":"v3.2-exp ships day-1 ascend/cann support","type":"reporting","url":"https://www.tomshardware.com/tech-industry/deepseek-new-model-supports-huawei-cann"},{"author":null,"id":"source-e8230989327d1388040d","note":"\"second deepseek moment\" — kimi k2","originalId":"src-ci-xpert-kimi","published":{"precision":"day","text":"2025-07-21"},"publisher":"xpert.digital","title":"second deepseek moment — kimi k2","type":"reference","url":"https://xpert.digital/en/sputnik-the-moment-of-artificial-intelligence/"},{"author":null,"id":"source-def3a541e976061d5949","note":"2025 return 56.55% (#2); >rmb 70b aum; 3/5-yr means","originalId":"src-hp-21jingji-hf2025","published":{"precision":"day","text":"2026-01-14"},"publisher":"21jingji.com","title":"high-flyer 2025: +56.55%, >rmb 70b","type":"reporting","url":"https://www.21jingji.com/article/20260114/herald/c28662157ae4f0d9f2c7c944101a35ad.html"},{"author":null,"id":"source-f99d11cc91e37d3d6dc1","note":"official app announcement","originalId":"src-hp-app-news","published":{"precision":"day","text":"2025-01-15"},"publisher":"deepseek.com","title":"official app announcement","type":"primary","url":"https://www.deepseek.com/en/news/deepseek-app/"},{"author":null,"id":"source-7c345f5754ef4659dd7a","note":">rmb 50b; hedged products to zero","originalId":"src-hp-bloomberg-neutral","published":{"precision":"day","text":"2024-10-21"},"publisher":"bloomberg","title":"top china quant winds down strategy","type":"reporting","url":"https://www.bloomberg.com/news/articles/2024-10-21/top-china-quant-winds-down-strategy-pummeled-by-market-rally"},{"author":null,"id":"source-a17c609e873002bfa73c","note":"v4 delayed for ascend-tuned stack (yuyuantantian)","originalId":"src-hp-bloomberg-v4delay","published":{"precision":"day","text":"2026-04-26"},"publisher":"bloomberg","title":"v4 delayed for ascend-tuned stack (yuyuantantian)","type":"reporting","url":"https://www.bloomberg.com/news/articles/2026-04-26/deepseek-v4-delay-shows-shift-to-china-chips-cctv-account-says"},{"author":null,"id":"source-27c1e329da17603af0fd","note":"may-8 round launch; liang's rmb 20b; ff1/ff2 cost figures; ~84% control","originalId":"src-hp-coinlive-funding","published":{"precision":"month","text":"2026-05"},"publisher":"coinlive.com","title":"may-8 round launch","type":"reporting","url":"https://www.coinlive.com/news/deepseek-s-liang-wenfeng-we-don-t-want-alibaba-s-money-we-ll-contribute"},{"author":null,"id":"source-298530916a213e82b46c","note":"math-7b (arxiv 2402.03300); grpo; author list","originalId":"src-hp-deepseekmath-github","published":{"precision":"day","text":"2024-02-05"},"publisher":"github","title":"the deepseek-math repo — grpo","type":"repo","url":"https://github.com/deepseek-ai/deepseek-math/"},{"author":null,"id":"source-33b6ba38ceadb66b8b98","note":"flashmla; v3.2 sparse-kernel updates","originalId":"src-hp-flashmla","published":{"precision":"month","text":"2025-02"},"publisher":"github","title":"the flashmla repo","type":"repo","url":"https://github.com/deepseek-ai/FlashMLA"},{"author":null,"id":"source-904e7f9caba419702302","note":"gtc-2022 blog: 10k-a100 cluster, ≥90% util, 3fs/hfreduce/hfai.nn, 50+ labs, 2.7m gpu-hrs","originalId":"src-hp-hfai-blog","published":{"precision":"year","text":"2022"},"publisher":"high-flyer.cn","title":"the hfai blog: the 10k-a100 cluster","type":"primary","url":"https://www.high-flyer.cn/blog/hfai/"},{"author":null,"id":"source-00a84e4dd575df716128","note":"official timeline 2008–2022; ff1/ff2 specs+costs; donations; hq address 汇金国际大厦a座14层","originalId":"src-hp-highflyer-history","published":{"precision":"month","text":"2026-09"},"publisher":"high-flyer.cn","title":"official timeline 2008–2022","type":"primary","url":"https://www.high-flyer.cn/history/"},{"author":null,"id":"source-afb542d09f0a4ec6ed22","note":"fire-flyer ii stats (8tb/s read, feb-2022 stats)","originalId":"src-hp-highflyer-home","published":{"precision":"month","text":"2026-09"},"publisher":"high-flyer.cn","title":"fire-flyer ii stats (8tb/s read, feb-2022 stats)","type":"primary","url":"https://www.high-flyer.cn/en/"},{"author":null,"id":"source-3c4d06b451215cd65fb0","note":"r2 status: unannounced, unreleased as of jul-2026; delay reporting recap","originalId":"src-hp-layer3-r2","published":{"precision":"month","text":"2026-07"},"publisher":"layer3labs.io","title":"r2 status: unannounced, unreleased as of jul-2026","type":"reporting","url":"https://www.layer3labs.io/guides/deepseek-r2-explained"},{"author":null,"id":"source-83745b9a03c1c5e17116","note":"math-v2: imo-25 & cmo-24 gold-level; putnam 118/120","originalId":"src-hp-mathv2","published":{"precision":"day","text":"2025-11-27"},"publisher":"arxiv.org","title":"math-v2: imo-25 & cmo-24 gold-level","type":"primary","url":"https://arxiv.org/html/2511.22570"},{"author":null,"id":"source-41500b9a68e5088f3bfa","note":"authors: shao, luo, lu, ren, hu, ye, gou, ma, zhang","originalId":"src-hp-mathv2-github","published":{"precision":"day","text":"2025-11-27"},"publisher":"github","title":"authors: shao, luo, lu, ren, hu, ye, gou, ma, zhang","type":"repo","url":"https://github.com/deepseek-ai/DeepSeek-Math-V2"},{"author":null,"id":"source-da084182120e71f10ea5","note":"app released jan 10 on ios+android","originalId":"src-hp-mobiledevmemo","published":{"precision":"day","text":"2025-01-29"},"publisher":"mobiledevmemo.com","title":"app released jan 10 on ios+android","type":"reporting","url":"https://mobiledevmemo.com/deepseek-and-consumer-ais-first-mover-disadvantage/"},{"author":null,"id":"source-84d7b54e8e39bb80541e","note":"neutral exit details; sept-11 earlier notice; <5% aum","originalId":"src-hp-nbd-neutral","published":{"precision":"day","text":"2024-10-19"},"publisher":"m.nbd.com.cn","title":"neutral exit details","type":"reporting","url":"https://m.nbd.com.cn/articles/2024-10-19/3596588.html"},{"author":null,"id":"source-7a9748e200d02457d23f","note":"open source week index; daily drops list","originalId":"src-hp-openinfra","published":{"precision":"day","text":"2025-02-24"},"publisher":"github","title":"open source week index","type":"repo","url":"https://github.com/deepseek-ai/open-infra-index"},{"author":null,"id":"source-ba66c76fcb42fa01110c","note":"prover-v2 (7b/671b), lean 4","originalId":"src-hp-prover-github","published":{"precision":"day","text":"2025-04-30"},"publisher":"github","title":"prover-v2 (7b/671b), lean 4","type":"repo","url":"https://github.com/DeepSeek-AI/DeepSeek-Prover-V2"},{"author":null,"id":"source-83afeaedd663fa1fd697","note":"huawei onsite team; ascend→inference-only; no fp8","originalId":"src-hp-register-huawei","published":{"precision":"day","text":"2025-08-14"},"publisher":"theregister.com","title":"huawei onsite team","type":"reporting","url":"https://www.theregister.com/software/2025/08/14/dodgy_huawei_chips_nearly_sunk_deepseeks_next_gen_r2_mode/753616"},{"author":null,"id":"source-2ebfdd7106ca4b48e28d","note":"ft: ascend training failure delayed r2","originalId":"src-hp-reuters-r2ft","published":{"precision":"day","text":"2025-08-14"},"publisher":"reuters","title":"ft: ascend training failure delayed r2","type":"reporting","url":"https://www.reuters.com/world/china/deepseeks-launch-new-ai-model-delayed-by-huawei-chip-issues-ft-reports-2025-08-14/"},{"author":null,"id":"source-11d00c40b93f5574acad","note":"r2 stalled; liang unsatisfied; planned may","originalId":"src-hp-reuters-r2june","published":{"precision":"day","text":"2025-06-26"},"publisher":"reuters","title":"deepseek-r2 launch stalled; ceo balks","type":"reporting","url":"https://www.reuters.com/world/china/deepseek-r2-launch-stalled-ceo-balks-progress-information-reports-2025-06-26/"},{"author":null,"id":"source-c5488a0485bbd430b7b9","note":"#1 us app store jan 26; 51 countries (appfigures)","originalId":"src-hp-techcrunch-appstore","published":{"precision":"day","text":"2025-01-27"},"publisher":"techcrunch","title":"#1 us app store jan 26","type":"reporting","url":"https://techcrunch.com/2025/01/27/deepseek-displaces-chatgpt-as-the-app-stores-top-app/"},{"author":null,"id":"source-4e9f90a2f630aa66eaf8","note":"italy/taiwan/navy/texas ban roundup","originalId":"src-hp-techcrunch-bans","published":{"precision":"day","text":"2025-02-03"},"publisher":"techcrunch","title":"italy/taiwan/navy/texas ban roundup","type":"reporting","url":"https://techcrunch.com/2025/02/03/deepseek-the-countries-and-agencies-that-have-banned-the-ai-companys-tech/"},{"author":null,"id":"source-446a0de49e666c3afad0","note":"17 jul 2023 founding; 160 staff; release table; anthropic note","originalId":"src-hp-wiki-deepseek","published":{"precision":"month","text":"2026-09"},"publisher":"wikipedia","title":"17 jul 2023 founding","type":"reference","url":"https://en.wikipedia.org/wiki/DeepSeek"}]}
````