500亿 · 梁文锋自掏200亿 · V4.1六月见 — 不为资本低头，为理想备好弹药，一场21天估值翻5倍的故事
PART 1
梁文锋，1985年生于广东湛江，浙江大学本硕毕业。他的故事从金融圈开始：2008年，刚毕业的他带着7人团队，用机器学习模型做量化交易，三个月实现500%收益。
2015年，幻方量化正式成立。2021年管理规模突破千亿，2025年平均收益率达56.6%，仅当年就为梁文锋带来超7亿美元收入。[4] 同年，他以115亿美元身家位列福布斯中国富豪榜第34位。[5]
但梁文锋的野心不止于金融。
2020年3月，幻方量化做了一个当时看来极为超前且重资产的决策——投资上亿元自建GPU计算集群。"萤火一号"搭载上千张高端GPU投入运行。这一远超行业常规做法的投入，为后来AGI布局埋下了决定性伏笔。[4]
2023年7月，杭州深度求索注册成立，DeepSeek正式亮相。梁文锋立下三条铁律：不融资、不上市、不商业化。
"DeepSeek的目标是做世界顶级的通用大模型，不是为了赚钱，也不是为了上市。资本会追求短期回报，商业化会妥协技术路线——这些都会干扰我们的核心目标。"[4]
在资本疯狂涌入AI的2023-2024年，这是一个近乎"叛逆"的选择。澎湃新闻评价，DeepSeek不是AI圈的正规军，而是"独树一帜的镇元子"——技术硬、性子倔、不站队。
而技术的回报，超出了所有人的预期：
2024年5月 — DeepSeek-V2发布，创新MoE架构震动行业。
2024年12月 — DeepSeek-V3开源，53页技术细节全公开。
2025年1月 — DeepSeek-R1在数学、代码、推理上比肩OpenAI o1。论文登上英国《自然》杂志，训练成本仅29.4万美元。[6]
"三不铁律"一度成为行业美谈——原来不靠融资、不靠商业化，也能做出世界一流的模型。
Born in 1985 in Zhanjiang, Guangdong, Liang Wenfeng holds degrees from Zhejiang University. He started in quantitative finance: in 2008, his 7-person team used machine learning for quant trading, achieving 500% returns in three months.
He founded High-Flyer Quant in 2015, which grew to manage over 100 billion yuan by 2021. In 2025, High-Flyer delivered a 56.6% average return, netting Liang over $700 million that year alone.[4] He made the Forbes China Rich List with an $11.5 billion fortune, ranking 34th.[5]
But Liang's ambition went beyond finance.
In March 2020, High-Flyer made a highly unconventional decision — pouring profits into building its own GPU cluster."Firefly 1" carried thousands of high-end GPUs, laying the groundwork for what would become DeepSeek.[4]
In July 2023, DeepSeek was officially founded. Liang established three iron rules: no outside funding, no IPO, and no commercialization.
"DeepSeek's goal is to build world-class general-purpose AI models — not to make money, not to go public. Capital chases short-term returns; commercialization compromises technical direction."[4]
The Paper (澎湃新闻) called DeepSeek "not the celestial army, but a maverick Zhenyuanzi" — technically strong, stubbornly independent, refusing to pick sides.
Then came the technical payoff:
May 2024 — DeepSeek-V2 launched with an innovative MoE architecture.
Dec 2024 — DeepSeek-V3 open-sourced with a 53-page technical report.
Jan 2025 — DeepSeek-R1 matched OpenAI o1 on math, code, and reasoning. Published in Nature. Training cost: just $294,000.[6]
PART 2
从4月初到5月上旬，DeepSeek的估值经历了戏剧性的"三级跳"：
From early April to early May, DeepSeek's valuation went through a dramatic triple jump:
数据来源：The Information、36氪、新浪财经、天眼查、澎湃新闻等。DeepSeek官方尚未置评。
PART 3
🔹 算力：从"算法巧胜"到"消耗战"
两年前的大模型竞赛还可以靠算法巧思"四两拨千斤"，如今则是赤裸裸的算力消耗战。V4系列1.6T参数、V4.1即将多模态——每一步都需要天文数字的算力投入。36氪说得直白："幻方量化再有钱，也撑不住一场与全球巨头正面交锋的算力军备竞赛。"[4]
🔹 人才：明星研究员接连出走
自2025年下半年以来，至少5名核心研发人员出走：
• 郭达雅 → 字节跳动，Agent负责人之一
• 罗福莉 → 小米，雷军亲自挖角
• 王炳宣、魏浩然等也陆续离开
• Meta开出的天价合同：4年2亿至3亿美元
此次融资的重要目的之一，就是给员工做期权定价，用股权留住人才。
🔹 产品化：实验室必须走向市场
据The Information报道，DeepSeek员工已开始向企业客户推广模型。V4.1的MCP协议支持正是为了企业接入。一个实验室可以只关心benchmark。一家AI公司必须关心客户、收入、交付和成本。
🔹 Compute: From Algorithmic Brilliance to a War of Attrition
Two years ago, LLM contests could be won with clever algorithms. Today, it's a raw compute war. As 36Kr put it bluntly: "High-Flyer has deep pockets, but not deep enough for a compute arms race against global giants."[4]
🔹 Talent: Star Researchers Are Leaving
Since H2 2025, at least five core researchers have left:
• Guo Daya → ByteDance as Agent lead
• Luo Fuli → Xiaomi, poached by Lei Jun
• Wang Bingxuan, Wei Haoran also departed
• Meta's offer: 4-year contracts worth $200-300M total
The funding round is partly about creating option pools to retain talent.
🔹 Productization: The Lab Must Go to Market
DeepSeek employees have begun pitching models to enterprise clients. V4.1's MCP protocol support is designed for enterprise integration. A lab can focus on benchmarks. A serious AI company must focus on customers, revenue, delivery, and cost.
PART 4
梁文锋自掏腰包，占募资总额40%，成为最大投资方
创始人自掏200亿——控制权保卫战
在全球AI史上，这种"创始人大额领投"的模式极为罕见。业内分析，200亿背后大概率有银行授信支持。[13] 如果梁文锋在这一轮只小额跟投甚至不参与，外部投资人的定价权会更大。200亿买下的，是在这个估值区间主导对话的权利。
只收"两种钱"
国家大基金洽谈领投——战略资本，不干预技术路线，提供算力、政策资源
腾讯拟出资约60亿元，获约2%股权——产业资本，生态协同
传统财务VC（红杉、高瓴等）全部被拒。阿里巴巴未能在条款上达成一致。[4]
先锁股权，再开门迎客
4月27日，梁文锋通过直接增资，将持股从1%提升至34%，叠加间接持股合计控制84.29%。无论外部资本进来多少，控制权都牢不可破。[7][8]
Liang personal investment — 40% of total raise, making him the largest investor
$2.8 Billion for What? Control.
A founder personally investing $2.8 billion is virtually unprecedented in global AI. Industry insiders told KCB that the capital likely involves bank credit lines.[13] The $2.8 billion buys the right to dominate conversation at this valuation.
Only Two Kinds of Money Accepted
China IC Fund (Big Fund) in talks to lead — strategic, won't interfere with technical direction
Tencent investing ~$850M for ~2% equity — industrial capital for ecosystem synergy
Traditional VCs (Sequoia, Hillhouse, etc.) all turned away. Alibaba failed to reach terms.[4]
Lock Down Ownership First, Then Open the Door
On April 27, Liang boosted his direct stake from 1% to 34%, giving him total effective control of approximately 84.29%. No matter how much external capital enters, Liang's control is locked in.[7][8]
PART 5
DeepSeek的融资不是孤例。2026年5月7日至14日，一周之内中国大模型公司轮番刷新估值：
Between May 7-14, Chinese LLM companies rewrote valuation records in a single week:
5月7日 — 月之暗面(Kimi)完成约$20亿D轮，估值破$200亿。
5月8日 — 阶跃星辰传出近$25亿融资，产业链资本集中入场。
5月13日 — 智谱港股大涨36.9%，市值突破5000亿港元；MiniMax涨18.46%。[13]
一组数据值得细品：智谱AI 2025年营收不足8亿元，市值超5000亿港元；科大讯飞年营收271亿元、净利8.39亿元，市值约1187亿元。亏损中的智谱，估值是盈利中的科大讯飞近4倍。
资本来源也在变化——从"美元+互联网"到"国资+产业链资本"。大模型已不只是商业故事，更是国家科技竞争力的关键筹码。[13]
A revealing comparison: Zhipu AI's 2025 revenue was under RMB 800M, yet its market cap surpassed HK$500B. iFlytek, with RMB 27.1B in revenue, had a market cap of roughly RMB 118.7B. A loss-making Zhipu was valued at nearly 4x profitable iFlytek.
The shift in capital sources from "US dollar + internet" to "state capital + supply chain" tells a deeper story: LLMs are no longer just a commercial bet — they are a critical lever for national tech competitiveness.[13]
PART 6
据The Information报道，DeepSeek计划2026年6月推出V4.1。三大升级：
V4.1 is scheduled for June 2026, bringing three upgrades:
🖼️ 全模态覆盖 — 新增图像、音频理解
🔗 原生MCP协议 — 企业客户可直接接入真实流程
🧰 企业级工具链 — 配套发布
此外，下半年将支持华为昇腾算力——从英伟达到国产芯片的适配已启动。[10] 旧模型名将在7月24日停用。
Also confirmed: Huawei Ascend compute support in H2 2026 — shift from NVIDIA to domestic chips has begun.[10] Legacy model names will be deprecated on July 24.
EPILOGUE
V4发布时，DeepSeek在末尾引用了一句古语：
"不诱于誉，不恐于诽，率道而行，端然正己。"
—— 荀子《非十二子》[10]
不被赞誉诱惑，不被诽谤吓倒。但"率道而行"的前提，是先有足够的弹药。
500亿是弹药。200亿是姿态。V4.1是下一步。
曾经坚守"三不"的DeepSeek，正在用一家公司的方式继续做AI。梁文锋的500亿不是妥协的代价，而是清醒的选择——理想主义者最需要保护的，从来不是"不拿钱"的姿态，而是能持续做出世界一流模型的能力。
在中国AI的赛场上，从不缺追逐风口的热钱。缺的是那些钱烧完、风口过去之后，还能留下来继续往前走的人。DeepSeek用"不融资"证明过技术本身的价值，现在又用"融对资"证明理想主义者也能玩好资本游戏。真正值得关注的，不是它拿了多少钱，而是它拿了钱之后要做什么——那才是这份坚持和清醒的真正意义。希望它能走得更远。
When V4 launched in April, DeepSeek ended its release note with an ancient Chinese proverb:
"Not seduced by praise, not intimidated by slander — follow the Way and stand upright."
— Xunzi, "Against the Twelve Masters"[10]
But "following the Way" requires ammunition.
$7.3 billion is ammunition. $2.8 billion is a stance. V4.1 is the next step.
DeepSeek — the lab that once rejected venture capital — is learning to operate as a company. Whether it can complete this transformation without losing its soul may be the most compelling story in Chinese AI for 2026.
Idealism isn't wrong.
But idealism needs to be fed.
📚 参考资料
[1] The Information, "DeepSeek to Raise $7 Billion", 2026.05
[2] 新浪财经，《DeepSeek 500亿天价融资》，2026.05.11
[3] 凤凰科技/量子位，2026.05.09
[4] 36氪/澎湃新闻，《梁文锋的资本运作》，2026.05.13
[5] 福布斯中国内地富豪榜，2025
[6] Nature，DeepSeek团队论文，2025.01
[7] 天眼查工商信息，2026.04.27
[8] 搜狐财经，2026.04.28
[9] 上海证券报，引述渠道人士报道，2026.05
[10] 量子位/36氪，DeepSeek-V4发布报道，2026.04.24
[11] 百度百科"DeepSeek-V4.1"词条
[12] ChooseAI, "DeepSeek V4.1 Locked for June", 2026.05.09
[13] 科创板日报，《五强格局成型》，2026.05.15
[1] The Information, "DeepSeek to Raise $7 Billion", May 2026
[2] Sina Finance, May 11, 2026
[3] Phoenix Tech/QBitAI, May 9, 2026
[4] 36Kr / The Paper, May 13, 2026
[5] Forbes China Rich List, 2025
[6] Nature, DeepSeek team paper, Jan 2025
[7] Tianyancha, Apr 27, 2026
[8] Sohu Finance, Apr 28, 2026
[9] Shanghai Securities News, May 2026
[10] QBitAI/36Kr, DeepSeek-V4 launch, Apr 24, 2026
[11] Baidu Baike "DeepSeek-V4.1"
[12] ChooseAI, "DeepSeek V4.1 Locked for June", May 9, 2026
[13] KCB Daily, "Five-Headed Pattern Forms", May 15, 2026