hraness

side quests

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

the founders' quests

  1. 2019–an ordinary little pig · continuingthe 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.high-flyer.cn, 2026-09

the lab's quests

  1. 2021–the idle-time compute program · continuingthe 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.high-flyer.cn, 2026-09high-flyer.cn, 2022high-flyer.cn, 2026-09
  2. 2025-02-24 – 2025-03-01open source week · endedopen 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.github, 2025-02-24github, 2025-02scmp.com, 2025-03-01gigazine.net, 2025-03-03
  3. 2025-01–the consumer app · continuingthe 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.deepseek.com, 2025-01-15mobiledevmemo.com, 2025-01-29techcrunch, 2025-01-27techcrunch, 2025-02-03scmp.com, 2025-09
  4. 2025-04–the ascend hedge · continuingthe 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.theregister.com, 2025-08-14reuters, 2025-08-14theregister.com, 2025-08-22tomshardware.com, 2025-09-29bloomberg, 2026-04-26
  5. 2025-05 – 2026r2 · endedthe 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.reuters, 2025-06-26reuters, 2025-08-14the-decoder.com, 2025-08layer3labs.io, 2026-07
  6. 2026-05–the ipo track · activethe 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.hindustantimes.com, 2026-09coinlive.com, 2026-05
  7. 2024-10–the fund keeps earning · continuingthe 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.21jingji.com, 2026-01-14bloomberg, 2024-10-21m.nbd.com.cn, 2024-10-19

the bench's quests

  1. 2024-02–the algorithm that escaped · handed-offgrpo — 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.github, 2024-02-05linkedin, 2025-11-07xpert.digital, 2025-07-21
  2. 2024-04–prover · continuingthe theorem-proving line: prover-v1.5, then prover-v2 in april 2025 — a 7b worker plus the 671b teacher for lean 4 proofs.github, 2025-04-30
  3. 2024-12–vl2 and janus · continuingthe 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.wikipedia, 2026-09sydney.edu.au, 2025-01-29
  4. 2025-11–math-v2 · continuingself-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.arxiv.org, 2025-11-27github, 2025-11-27

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