people
liang wenfeng and the people around both companies — the zju classmates, the fund's partners, the research bench, and the diaspora.
the founder
- liang wenfeng 梁文锋 · founderfounder and ceo of both companies — the lab and the fund. b. 1985 in mililing village, wuchuan, zhanjiang, guangdong; father a primary-school teacher. entered zhejiang university at 17, meng in information and communication engineering under xiang zhiyu (thesis: low-cost ptz-camera target tracking). the 2008–2015 automated-trading years are his own telling — rmb 80,000 to nine-figure-rmb wealth, self-reported. per the 36kr profiles he still 'reads papers, writes code, and joins group discussions' like any researcher; he is a co-author on the technical papers — wenfeng liang on the fire-flyer ai-hpc paper and the v3/r1 author lists — unusual for a ceo. near-zero public profile until january 20, 2025, when he briefed premier li qiang the same day r1 shipped; february 17 he sat in xi jinping's private-enterprise symposium; april 17 he met jensen huang about compliant chip designs. wealth: hurun debut rmb 33b, march 2025; bloomberg ~$36b in july 2026 — the world's richest ai-model founder. donated rmb 138m to charity in 2022 under the pseudonym 'an ordinary little pig.'
- xu jin 徐进 · co-founderco-founder since jacobi (2013). zju chu kochen college mixed-honors class, then a phd in signal and information processing — a main member of the yutu lunar-rover visual-navigation research project. co-founded 每日科技 (now getui/merit interactive, a listed data company), then huawei's shanghai research institute. long the technical director of the quant entities and legal representative of zhejiang jiuzhang; ~7.5% of jiuzhang.
- zheng dawei 郑达韡 · co-founderzju classmate and named partner — a researcher at shanda innovation institute in 2011–12 before jacobi. co-leads the ai side of high-flyer with li huan; ~7.5% of jiuzhang.
the writing
liang wenfeng's public corpus is thin by design — one forum speech, two 暗涌 interviews, and the technical reports he co-authors. the papers are the lab's voice; the release notes say the rest.
- 2024-08fire-flyer ai-hpc · document, arxiv / sc24 · the researchersthe fire-flyer ai-hpc paper — how ~10,000 pcie a100s approximate dgx-a100 performance at ~half the cost and 60% of the energy, with the hfreduce comms library and 3fs filesystem. the ~50-author list is signed 'deepseek ai' and includes wenfeng liang — documentary proof the fund's infrastructure team became the lab.sc24 technical program
- 2024-12-26deepseek-v3 technical report · document, arxiv · the researchersthe deepseek-v3 technical report — 671b/37b-active moe, 14.8t tokens, and the claim that launched a market event: 2.788m h800 gpu-hours ≈ $5.576m, explicitly scoped to the final official pre-training run and excluding prior research and ablations.arxiv 2412.19437; the $5.6m figure misattributed to r1 for months
- 2025-02-24open source week · post, github · the researchersthe open-infra-index — open source week's six daily drops: flashmla, deepep, deepgemm, dualpipe, eplb with profile data, then 3fs and smallpond. the fund's internal stack — hfreduce, hfai.nn, the fire-flyer filesystem — published as public goods.six repos in six days
- 2025-09-17deepseek-r1 (nature 645:633–638) · document, nature · the researchersguo, yang, zhang et al., 'deepseek-r1' — the first major llm through full peer review. the nature version adds the disclosures the preprint lacked: ~$294k of rl compute atop the base model, and the statement that r1 did not train on openai-generated reasoning examples while conceding the base model ingested web ai-content like all llms.submitted feb 14, accepted jul 17, published sep 17 — the peer-review precedent
- 2025-11-27deepseekmath-v2 · document, arxiv · the researchersdeepseekmath-v2 — self-verifiable theorem proving at gold-level imo-2025 and cmo-2024, 118/120 on putnam 2024. the proof line running on scaled test-time compute.arxiv 2511.22570
- 2026the '64 quotes' investor-call leak · document, geopolitechs (unverified) · liang wenfeng 梁文锋the purported 'investor call' transcript — '64 quotes' circulated on aggregator posts: ~20,000 'h-equivalent' gpus, '12–18 months behind the us,' api priced to recoup hardware in ~10 months, 'we don't really have an organization — we're driven by a vision.' unverified — likely leaked notes from the 2026 fundraising roadshow; cataloged with that flag.unverified — treated as a leak, not a transcript
the team
- chen zhe 陈哲named partner — a former guangfa securities asset-management quant who runs pricing: options and bonds.
- li huan 李欢named partner — co-leads the ai side of high-flyer with zheng dawei.
- lu zhengzhe 陆政哲the current ceo of high-flyer — the fund's public-facing executive while liang runs the lab. ex-china-merchants-system background per the august 2025 rebate-case reporting.
- yan wentao 严文韬the lab's first cfo, hired ~september 2025 — b. 1991, an ex-tencent invest / h capital / hillhouse ventures partner whose portfolio included minimax and zhipu. the hire was widely read as fundraising groundwork, seven months before the first external round opened.
- the researchersthe lab runs flat: no titles, no departments, bottom-up staffing, and anyone can call training-cluster gpus without approval; the hiring screen is 'passion and curiosity' plus competition medals, with pay reportedly benchmarked above bytedance offers. the papers name the names: guo daya 郭达雅 (b. 1994, sun yat-sen phd and msra — codebert/graphcodebert — core contributor to coder, math, prover, v2, coder-v2, v3, and r1, and co-creator of grpo, >37k citations); zhihong shao, lead on deepseekmath, the prover line, and math-v2; plus the recurring core across the v2/v3/r1/fire-flyer author lists — damai dai (moe), aixin liu, junxiao song, chengqi deng (infra chief per reporting), peiyi wang, runxin xu, qihao zhu, deli chen, zhenda xie, wangding zeng, xiao bi. no public cto or chief-scientist title exists; release notes and papers are the lab's only consistent public voice.
- the diasporathe talent-war inversion — the lab famous for raising unknown graduates became the poaching ground once external money arrived. luo fuli, described as a key v3 architect, left for xiaomi's mimo as llm lead; wang bingxuan went to tencent hunyuan; chong ruan became deeproute.ai's chief scientist; wei haoran of the ocr work departed ~february 2026; and guo daya left in april 2026 for bytedance seed as an agent lead. ~10 names on the v4 report's ~300-author list are flagged former employees.
interviews and talks
- 2019-09一名程序员眼里中国量化投资的未来 · golden bull forumliang's rare public speech at the golden bull forum — 'the future of china's quant investing through a programmer's eyes' — predicting head-concentration of quant aum and multi-strategy superposition, two years before the fund hit its ~rmb 100b peak.
- 2023-05-24疯狂的幻方：一家隐形ai巨头的大模型之路 · 36kr / 暗涌'crazy high-flyer: an invisible ai giant's llm road' — the first-ever public interview, given to 36kr/暗涌's 于丽丽 weeks after the agi pivot. names 深度求索, discloses the gpu ladder, and explains why vcs were never an option: 'we want to do it, we can do it — that makes us one of the right people.'
- 2024-07-17揭秘deepseek：一个更极致的中国技术理想主义故事 · 36kr / 暗涌'a more extreme chinese techno-idealism story' — the canonical text, called one of the most-cited chinese ai texts globally. the cost-plus pricing doctrine, the contributor-not-free-rider argument, the bottom-up org with unbounded cluster access, and the agi timeline: '2, 5 or 10 years — within our lifetimes.'
the philosophy
themes tracked across the writing: technological idealism, open weights as strategy, constraint as engine, originality over imitation, the whale's money, young and homegrown, research, not products, the long bet.
- 2008–the self-reported originthe origin story is his own telling — the rmb 80,000 compounding >100% annually into nine-figure-rmb wealth comes from liang's 2015 retelling, relayed by bianews, with no independent verification. the birthdate is '1985' everywhere; the may-22 date is single-source. dji's wang tao reportedly tried to recruit him — reported, not confirmed.the dossier keeps the story and the flag: the self-made-trader account is load-bearing mythology until a second source exists.
- 2025-03the travel curbsthe information reported that some deepseek staff must surrender passports to high-flyer, that zhejiang officials screen would-be investors, and — per the wsj — that beijing warned top ai founders against us travel. anonymous-sourced reporting throughout; neither company has detailed a response.documented state proximity is real — the symposiums are confirmed; the degree of directive control is known only through anonymous-sourced reporting.
- 2024–2025the rebate scandalthe whale's compliance stain: high-flyer marketing director li cheng was detained in november 2024; reporting in august 2025 detailed a 2018–2023 scheme with a china merchants securities branch manager skimming ~rmb 118m of trading-commission rebates — over rmb 20m to li personally. cms executive gao xiang was investigated in may 2025.high-flyer's position: individual misconduct the company was unaware of. the case is the first real compliance mark on the patron.
- 2025–2026the poaching groundthe flat structure's cost: no titles, no departments, unbounded cluster access — and then the poaching began. ~10 v4-report authors now sit at bytedance seed, tencent hunyuan, deeproute, and xiaomi; guo daya's april 2026 exit for seed's agent lead was the headline departure.read both ways at once: validation of liang's 'we can build the top-50 talent ourselves' claim, and a real cost once external money arrived.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.