history
the company, sourced and dated — from the dorm-room trading bots to fire-flyer ii, the r1 shock, and the first raise.
the classmates2008–2015
liang wenfeng and his zhejiang university classmates start automating stock trades during the financial crisis, incorporate jacobi in 2013, and register the fund that becomes high-flyer in june 2015.
- 2008automated trading from the dormsliang wenfeng and zhejiang university classmates begin exploring fully-automated quantitative trading mid-financial-crisis — scraping market data from trading software with image processing and writing plug-in programs to crack trading interfaces. liang's 2015 retelling starts him with rmb 80,000 of personal capital compounding into nine-figure-rmb wealth — self-reported, unverified. the team is all-zju engineering; none have overseas or institutional trading experience.
- 2010the machine-vision master'sliang completes his meng at zju — advisor xiang zhiyu, thesis on low-cost ptz-camera target tracking — then moves to chengdu to experiment with ai applications, an amac registry listing him as a 'freelancer' until 2013. dji founder wang tao reportedly tried to recruit him on the strength of the machine-vision work.
- 2013jacobiliang and zju classmate xu jin incorporate hangzhou jacobi investment management — named for mathematician c.g.j. jacobi — the direct predecessor of high-flyer.
- 2015-03the first fundthe first fund product launches — 幻方沪深300指数增强1号, a csi-300 index-enhanced strategy.
- 2015-06-11high-flyer registeredhangzhou high-flyer technology is registered — later renamed zhejiang jiuzhang asset management ('nine chapters' honors the classical math text). the official history lists 2015 as the founding year; the brand 幻方 ('magic square') references the luoshu diagram. xu jin is the legal representative.
the ai fund2016–2021
high-flyer turns into an ai fund — the first deep-learning trade in 2016, all-ai strategies by 2017, a golden bull award, the first fire-flyer cluster, and a ~rmb 100b peak that ends in a public apology for a record drawdown.
- 2016-02-15the ningbo lpthe ningbo high-flyer quantitative investment management partnership is incorporated (amac p1032021) — the second regulated entity, and the reason some english sources give 'february 2016' as the founding date.
- 2016-10-21the first deep-learning tradethe first live stock position generated by a deep-learning model goes on the books, gpu-computed — per high-flyer's own timeline, which is the only account.
- 2017all-ai strategies'almost all' quant strategies run on ai models by year-end — self-reported on the official timeline.
- 2018the golden bullthe first golden bull award — china's top private-fund prize — and the company formally sets ai as its primary direction, beginning the hunt for large-scale compute.
- 2019the 百亿 fundaum passes rmb 10b ('百亿私募'); high-flyer capital management hong kong takes an sfc type-9 license; the 幻方ai research company is registered for 'ai algorithm and basic-application research.' liang gives a rare public speech at the golden bull forum — 'a quant has no fund managers; the fund manager is a rack of servers.'
- 2020fire-flyer ifire-flyer i enters service — ~rmb 200m and 1,100 accelerator cards on 200gbps interconnect. it retires about a year and a half later, a practice fleet.
- 2021the 千亿 peakthe portfolio peaks — reuters: 'totalled over 100 billion yuan by end of 2021' (~$14b). on 15 november the fund halts all new subscriptions; on 25 november it waives rmb-fund redemption fees.
- 2021-12-28the apologythe fund's first public crisis: a record 10.66% max drawdown across 497 products. an executive's wechat apology circulates on the 27th — 'we'll fully accept the scolding, just please don't hit us' — and the official statement follows on the 28th: 'the drawdown reached a historical maximum; we are deeply ashamed,' blaming long-horizon ai holdings and industry-wide strategy crowding.
the fleet2021–2023
the whale buys its fleet: ~10,000 nvidia a100s for fire-flyer ii, all before the october 2022 export controls — the decisive timing fact of the whole history. in april 2023 the fund announces an independent agi lab.
- 2021fire-flyer ii~rmb 1b goes into fire-flyer ii — ~10,000 nvidia a100 (pcie) gpus with a task-level time-sharing scheduler, the self-built 3fs file system, hfreduce, and hfai.nn. all bought before the october 2022 export controls; per 36kr, high-flyer was 'the only company outside the big tech platforms stockpiling 10,000 a100s.'
- 2022the cluster at workthe cluster's self-reported year: ≥96% average occupancy, 1.35m jobs, 56.74m gpu-hours — and ~27% of capacity donated as idle-time compute to 50+ university and institute labs.
- 2022-10-07the export controlsus export controls ban a100/h100 sales to china; the october 2023 rules close the a800/h800 workaround variants. every gpu in the fire-flyer ii fleet was already bought.
- 2022an ordinary little pigthe philanthropy year: the group donates rmb 221.38m across 15 charities, and liang personally gives rmb 138m under the pseudonym 一只平凡的小猪 — 'an ordinary little pig' — listed on high-flyer's own history page.
- 2023-04the agi pivothigh-flyer announces it will concentrate resources on agi in a new, independent research organization 'separate from finance,' quoting truffaut's advice to young directors: 'be desperately ambitious, and desperately sincere.' the date is mid-april — the 36kr interview gives the 11th; most coverage cites the 14th.
the lab2023–2024
deepseek incorporates in hangzhou and runs the release ladder — coder, llm, moe, math (grpo), v2's price war, the techno-idealism interview, and v3's 2.788m-gpu-hour final run.
- 2023-05深度求索the new org is named 深度求索 — deepseek, echoing qu yuan's 求索, 'to seek.' liang's first-ever interview runs in 36kr/暗涌: 'our llm work has nothing directly to do with quant or finance.'
- 2023-07-17incorporated in hangzhouhangzhou deepseek artificial intelligence basic technology research co., ltd. is incorporated — rmb 10m registered capital. liang holds 1% directly and ~99% sits with a ningbo limited partnership he controls through nested vehicles — ~84% effective ownership reported. no high-flyer entity appears on the cap table; the funding link is liang himself plus shared people and compute.
- 2023-11-02deepseek coderdeepseek coder — 1b to 33b parameters trained on 2t tokens — is the lab's first public model.
- 2023-11-29deepseek-llmdeepseek-llm 7b/67b — the dense base-model pair.
- 2024-01deepseekmoedeepseekmoe — the mixture-of-experts architecture line that v2 scales up.
- 2024-02-05deepseekmath and grpodeepseekmath 7b introduces grpo — group relative policy optimization, the rl algorithm that will later drive r1. the author list names zhihong shao, peiyi wang, and daya guo.
- 2024-05-06the price wardeepseek-v2 — 236b moe with mla and deepseekmoe — launches at ~rmb 1 per million input tokens and ignites china's llm price war: bytedance, alibaba, baidu, and tencent follow with cuts of 80–97%. semianalysis calls the paper possibly the year's best; jack clark calls the team 'a squad of unfathomable geniuses.' liang: 'we didn't mean to be a catfish; we just accidentally became one.'
- 2024-06-17coder-v2deepseek-coder-v2 — the code line catches up to the v2 architecture.
- 2024-07-17the techno-idealism interviewthe second 36kr/暗涌 interview — the canonical 'chinese techno-idealism' text: 'the real gap is originality vs. imitation'; 'open source is a cultural act more than a commercial one'; 'our problem has never been money — it's the embargo on high-end chips.'
- 2024-08the fire-flyer paperthe fire-flyer ai-hpc paper is published for sc24 — ~10,000 pcie a100s approximating dgx-a100 performance at ~half the cost and 60% of the energy. the ~50-author list is signed 'deepseek ai' and includes wenfeng liang: the fund's infrastructure team becoming the lab, on paper.
- 2024-09-05v2.5deepseek-v2.5 merges the chat and coder lines.
- 2024-10-19the neutral exithigh-flyer tells investors it will cut all hedged-product positions to zero and waive their management fees — exiting the market-neutral strategy after the a-share rally crushes it. bloomberg puts zhejiang high-flyer above rmb 50b at the time.
- 2024-11-20r1-lite-previewr1-lite-preview — the reasoning model's first public outing, api and web only.
- 2024-12-13vl2deepseek-vl2 — the vision-language line.
- 2024-12-26deepseek-v3deepseek-v3 — 671b total / 37b active moe, 14.8t tokens — ships open-weights with a technical report claiming 2.788m h800 gpu-hours ≈ $5.576m for the final pre-training run only, at an assumed $2/gpu-hour. silicon valley notices; the number the world will misread is now in print.
the shock2025-01 – 2025-03
r1 lands on inauguration day, the app passes chatgpt, nvidia loses ~$589b in a session, bans cascade across governments — and liang briefs the premier and attends xi's symposium in the same weeks.
- 2025-01-10the appthe deepseek assistant app hits the app store and google play (~jan 10; the official news post follows jan 15).
- 2025-01-20r1 and the premierdeepseek-r1 ships open-weights under mit — plus r1-zero and six distilled models. the same day, liang attends premier li qiang's symposium on the draft government work report as one of nine speakers — the split-screen that defined the week.
- 2025-01-24the sputnik postandreessen posts: 'deepseek r1 is one of the most amazing and impressive breakthroughs i've ever seen — and as open source, a profound gift to the world' — and calls it 'ai's sputnik moment.'
- 2025-01-26#1 in the app storethe deepseek app passes chatgpt to #1 free app on the us app store — and tops the charts in 51 other countries per appfigures.
- 2025-01-27the crashthe market shock: nvidia −17%, ~$589–600b of market cap — the largest single-day loss for a us company on record; broadcom −17%; nasdaq −3.1%; ~$1t wiped overall. trump calls it 'a wake-up call.' deepseek limits new registrations, citing 'large-scale malicious attacks.'
- 2025-01-28janus into the chaosjanus-pro-7b — the multimodal model — ships into the chaos. on the 29th, wiz discloses an exposed, unauthenticated deepseek clickhouse database holding chat history, api secrets, and backend details; it is secured after disclosure.
- 2025-01-29the distillation claimopenai tells the ft it has evidence linking deepseek to distillation of its models; microsoft reportedly flagged suspicious api exfiltration via openai developer accounts in late 2024; david sacks claims 'substantial evidence.' no public evidence drop ever follows — the claim stands as reported accusation, not demonstrated finding.
- 2025-01-30the ban cascadethe ban cascade opens: italy's garante orders deepseek blocked — the first country — calling its response 'totally insufficient.' the us navy and nasa follow; texas takes the first state ban on the 31st; australia, taiwan, and korean ministries follow into february.
- 2025-01-31the commerce probereuters reports us commerce is probing whether deepseek used restricted chips; organized smuggling is tracked through malaysia, singapore, and the uae.
- 2025-02-17the xi symposiumliang attends xi jinping's closed-door private-enterprise symposium — seated near the front alongside jack ma, pony ma, ren zhengfei, wang chuanfu, and lei jun; attendance is state-media confirmed via yuyuantantian. 'deepseek fever' sweeps china through february and march — telecoms, hospitals, and local governments deploy r1; tencent builds it into wechat search.
- 2025-02-24open source weekopen source week — six daily drops: flashmla, deepep, deepgemm, dualpipe and eplb with profile data, then 3fs and smallpond. the fund's internal infrastructure stack becomes public goods.
- 2025-02-27the 545% figurethe lab closes the week by disclosing a theoretical 545% daily profit margin on v3/r1 inference — flagged 'purely theoretical' in its own post, quoted without the caveat for months after.
- 2025-02the lobby pilgrimagethe huijin international mansion lobby — deepseek's registered address, in the same complex as high-flyer — becomes a pilgrimage site: 40–50 visitor batches a day at peak, foreign tourists included. the company declines all interviews and lets the papers speak.
the scrutiny2025
the year of scrutiny and normalization — passports held, r2 stalled on a failed ascend run, the r1 paper through nature peer review, the rebate scandal on the fund side, and the first poached researchers.
- 2025-03the passportsthe information reports some deepseek staff must surrender passports to high-flyer; zhejiang officials screen would-be investors; the wsj reports beijing warned top ai founders against us travel.
- 2025-03-24v3-0324deepseek-v3-0324 ships under mit — the quiet checkpoint refresh.
- 2025-03-27the rich listliang debuts on the hurun global rich list at rmb 33b (~$4.6b) — >80% of deepseek per hurun's accounting.
- 2025-04-09the h20 licensethe us requires licenses for h20 sales to china — the last compliant channel closes. on april 17 jensen huang flies to beijing and meets liang to discuss compliant next-gen chip designs, per the ft.
- 2025-04-30prover-v2deepseek-prover-v2 — 7b plus the 671b teacher — for lean 4 theorem proving.
- 2025-05-28r1-0528r1-0528 — stronger reasoning, fewer hallucinations, json and function calling. zvi's post-mortem: it 'did not have a moment.'
- 2025-06-26r2 stallsthe information via reuters: r2 is stalled — liang unsatisfied with performance — and the nvidia h20 shortage clouds the rollout.
- 2025-08-07the rebate scandalthe fund's governance stain: caixin-linked reporting details a 2018–2023 scheme in which high-flyer marketing director li cheng and a china merchants securities branch manager skimmed ~rmb 118m of trading-commission rebates through a fake-broker arrangement — over rmb 20m to li personally. high-flyer says it was individual misconduct the company was unaware of.
- 2025-08-14the ascend failurethe ft via reuters: r2 was delayed after a failed training run on huawei ascend — huawei sent an engineer team onsite and the run still failed. deepseek returns to nvidia for training and keeps ascend for inference only.
- 2025-08-21v3.1deepseek-v3.1 — hybrid think/non-think modes and ~840b additional training tokens. a wechat comment noting its ue8m0-fp8 format is 'for next-generation domestic chips' rallies chinese chip stocks.
- 2025-09the first cfothe lab hires its first cfo — yan wentao, b.1991, a former hillhouse ventures partner whose portfolio included minimax and zhipu — widely read as fundraising groundwork.
- 2025-09-17the nature paperthe r1 paper is published in nature (645:633–638; submitted feb 14, accepted jul 17) — the first major llm to pass full peer review. the review exchange surfaces two disclosures: r1's own rl stage cost ~$294k atop the base model, and deepseek's statement that r1 did not train on openai-generated reasoning examples.
- 2025-09-29v3.2-exp and dsav3.2-exp debuts deepseek sparse attention — and ships day-1 huawei ascend/cann and vllm support alongside it.
- 2025-11-27math-v2deepseekmath-v2 — self-verifiable theorem proving — posts gold-level imo-2025 and cmo-2024 results and 118/120 on putnam 2024 with scaled test-time compute.
- 2025-12-01v3.2deepseek-v3.2 general release, plus the v3.2-speciale heavy-reasoning endpoint.
- 2025-12the poaching groundthe talent-war inversion: the lab that raised unknown graduates becomes the poaching ground — luo fuli (a key v3 architect) to xiaomi mimo, wang bingxuan to tencent hunyuan, chong ruan to deeproute.ai as chief scientist; ~10 names on the v4 report's ~300-author list are flagged former employees.
the incumbent2026–
the disruptor behaves like an incumbent: the first external raise (~$7.4b at ~$52b reported), v4's preview and ga, api price hikes, ipo prep — and r2 never shipped.
- 2026-01-14the whale keeps earningthe whale keeps earning: high-flyer's 2025 books close at +56.55% mean product return — #2 among >rmb-10b chinese quant privates — on >rmb 70b aum.
- 2026-02-23the anthropic disclosureanthropic publishes its own account of 'industrial-scale' distillation campaigns — ~24k fraudulent accounts and >16m claude exchanges across deepseek, moonshot, and minimax; deepseek's ~150k+ exchanges are called the most technically sophisticated, including prompts to generate 'censorship-safe' answers on sensitive topics. an openai letter to lawmakers earlier in february alleges ongoing distillation. deepseek has not published a detailed rebuttal.
- 2026-04-24the v4 previewthe v4 preview: v4-pro (1.6t total / 49b active) and v4-flash (284b / 13b) with 1m-token context under mit. two days later bloomberg reports the launch was delayed months for deep huawei-ascend optimization, aligning with beijing's chip self-reliance push.
- 2026-05-08the first raisethe first external funding round ever opens: a ~rmb 300b (~$45b) valuation raising up to rmb 50b, with liang personally injecting up to rmb 20b. alibaba's talks collapsed over ecosystem-binding terms; tencent's proposed ~20% stake was declined; state capital — the national ic fund — is reported the likely lead. the purpose per liang: compute and r&d money, plus a market valuation anchor to retain talent.
- 2026-06the round closesthe round reportedly closes: ~$7.4b at a ~$52b valuation per forbes; linkedin lists the round june 1.
- 2026-07-14richest ai founderthe bloomberg billionaires index puts liang at ~$36b — the world's richest ai-model founder, above amodei and brockman.
- 2026-08-13v4-pro ga and the price hikev4-pro goes ga on app, web, and api with peak/off-peak pricing (off-peak at half price). a 'significant' api price increase lands days later — up to ~12x on cache hits. v4.1-flash follows on september 10 — natively multimodal, the smallest of the new architecture family, with v4-pro being phased out in its favor.
- 2026-09r2 never shippedas of this writing, r2 has never shipped — never even officially announced. the v-line is the flagship; whether the r-line was folded into it is unknown.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.