The TeardownMiniMax (稀宇科技)
All studies ▾
A case study · as of June 7, 2026

MiniMax: the fastest sprint from founding to an AI IPO

An independent, fully-cited, deliberately neutral teardown of MiniMax (稀宇科技) — how a Chinese AI 'six tiger' built an overseas consumer franchise and top-ranked low-cost open models, and the losses, price war and legal questions that define its next chapter.

HKEX: 0010033 sources · ~45% ChineseNeutral · evidence on both sides

In a little over four years MiniMax went from a few ex-SenseTime researchers to a publicly traded AI company — among the fastest founding-to-IPO runs anywhere[3]. It did it with an unusual recipe for a Chinese lab: sell to the world, not just China, and win on efficiency rather than raw spend. The open question is whether that recipe reaches profit before the price war, litigation and geopolitics catch up.

As of June 7, 2026, MiniMax is listed on the HKEX (00100) after a Jan 9, 2026 debut that surged 70%+ to a peak market cap near $11.5B[1]. For the first nine months of 2025 it reported $53.4M revenue (+~175%), 73% of it overseas, against a GAAP net loss of $512M (mostly non-cash fair-value charges; adjusted loss ~$186M)[2][25]. It serves 236M cumulative users[27] and ships top-ranked open models[18]. This site lays out the bull and bear case on each big question and leaves the verdict to you.

$53.4M
revenue, 9M 2025 (+~175% YoY)
73% from overseas [2]
$512M
GAAP net loss, 9M 2025
adj. loss ~$186M [25]
27.6M
monthly active users (Sep 2025)
from 3.1M in 2023 [27]
~$11.5B
peak market cap on debut
HKEX, Jan 9 2026 [1]

Revenue is compounding off a tiny base

MiniMax's disclosed revenue rose from just $3.46M in 2023 to $30.5M in 2024 (+782%) and roughly $79M for full-year 2025 (+159%)[24]. The growth is real and overseas-led, but the absolute numbers remain small for an ~$11B company — the valuation prices in the franchise and the models, not today's revenue.

MiniMax revenue (US$M, disclosed)
202320249M'25FY'25

Revenue per MiniMax's IPO prospectus as reported by 36Kr and Wall Street CN[24][2]. 9M figure is Jan–Sep 2025; FY is full-year 2025.

The four questions this case study turns on

01

Can MiniMax turn 175% revenue growth into a profit?

Revenue hit $53.4M for the first nine months of 2025 (+~175%), and gross margin turned positive (23%). But the company still posted a $512M GAAP net loss — most of it non-cash fair-value charges, with an adjusted loss near $186M. Bulls see a clear path as costs fall; bears see a business that burned through a price war and isn't self-funding.

02

Is efficiency plus open weights a real moat?

MiniMax's bet is architectural — 'lightning'/linear attention plus MoE — to deliver near-frontier models cheaply, then release them open-weight (M1, M2). M2 was ranked the #1 open-weight model by Artificial Analysis. But open weights are easy to copy, and giving the model away (M2's API launched free) is a distribution win, not a revenue one.

03

Does overseas consumer reach set it apart from the other 'tigers'?

Unlike most Chinese rivals, MiniMax earns 73% of revenue overseas via consumer apps (Hailuo video, Talkie/Xingye companions) across 200+ countries. That global B2C franchise is unusual among Chinese AI labs — but it sits in front of a crowded field of six 'tigers', DeepSeek, and big-tech model arms all racing the same benchmarks.

04

Do the legal and geopolitical overhangs bite?

Disney, Universal and Warner sued over Hailuo's outputs (Sept 2025); Anthropic accused MiniMax of distilling Claude (Feb 2026); companion apps draw safety scrutiny; and US chip-export controls constrain compute. None has dented growth yet, but each shapes the cost, legality and trust the model depends on.

The balance of evidence, at a glance

Why the bull case holds

  • A genuinely global consumer franchise — 73% overseas revenue, 236M users across 200+ countries — that most Chinese rivals lack[15][27].
  • Top-ranked low-cost open models: M2 ranked #1 open-weight on Artificial Analysis; M1 trained its RL stage for $534,700[18][17].
  • Fast-rising, increasingly efficient economics: revenue +175%, gross margin from −24.7% to 23%+ as cloud costs fall[24][12].
  • Deep-pocketed backers (Alibaba ~13.7%, Tencent, miHoYo) and ~$1.05B cash post-raise for runway[22][26].

Why the bear case holds

  • Still deeply loss-making (adj. ~$186M in 9M2025) in a domestic price war that compresses everyone's margins[30][10].
  • Its core models are open-weight and quickly commoditized; M2's API launched free, pressuring direct monetization[19].
  • Unresolved litigation: a Disney/Universal/Warner copyright suit and an Anthropic 'distillation' accusation[28][4].
  • Companion-app safety scrutiny plus US chip-export controls that constrain compute and reach[29][31].
⚖️
What reasonable people disagree about: whether efficiency + open weights is a durable moat or a treadmill[18][19]; whether the overseas consumer lead converts to durable profit before the price war bites[15][10]; how much the GAAP loss overstates the real burn[25]; and whether the IP/distillation cases are noise or a structural threat to a model trained for global markets[28][4]. Each is genuinely contested in the sources.
🧭
This is an independent research compilation, not affiliated with MiniMax and not investment advice. MiniMax is a young public company; figures are point-in-time as of June 7, 2026 and several are from its IPO prospectus as reported by the press. Critical claims are attributed. A substantial share of sources are Chinese-language, with original text shown. See Methodology & Limitations for what may be wrong and Sources for the full bibliography.
Company & Timeline

Four years from a research idea to a listed AI company

MiniMax's arc: consumer apps first (Glow → Talkie/Xingye), then a multimodal model stack (Hailuo, abab, M-series), then one of the fastest founding-to-IPO runs in AI.

Founded Dec 2021Shanghai

MiniMax was a consumer-app company before it was a model company: Glow and Talkie/Xingye proved overseas demand, and the abab → M-series models followed. That sequence — distribution first, then frontier models — is why 73% of revenue is overseas and why the IPO came so fast.

From founding to IPO, the shortest path among AI companies.
original · zh ·从成立到IPO,历时最短的AI公司。
财联社 (Cailianpress) · on MiniMax passing its HKEX hearing · Dec 2025 · English is a translation from zh · source
  1. Dec 2021Founded in Shanghai

    Yan Junjie (闫俊杰, ex-SenseTime VP) and co-founders start MiniMax; the name nods to the minimax algorithm[5][7].

  2. Oct 2022Glow launches

    First consumer app — AI character chat — reaches ~5M users in months, then is pulled from Chinese app stores in 2023[5].

  3. 2023Talkie & Xingye

    Glow relaunches as Talkie internationally (Jun 2023) and 星野 (Xingye) in China (Sep 2023) — the companion-app franchise[5].

  4. Mar 2024Alibaba leads $600M

    A $600M round led by Alibaba values MiniMax at $2.5B; abab 6.5 MoE model ships in April[6].

  5. Mar 2024Hailuo AI (海螺)

    Launches Hailuo AI; the video-01 text-to-video model follows (Sep 2024), building the multimodal product line[5].

  6. Jan 2025MiniMax-01

    General text + visual models released; Speech-02 (30+ languages) follows in April[6].

  7. Jun 16, 2025M1 open-weight

    Releases M1 — the first open-weight large-scale hybrid-attention reasoning model, 1M-token context, RL stage trained for $534,700[17].

  8. Jul 2025~$4B valuation

    Reported to close a new round; ~$1.5B raised across 7 rounds at a ~$4B valuation; Alibaba holds ~13.7%[22].

  9. Oct 2025M2 open-weight (MIT)

    Ships M2 (230B/10B-active MoE) for coding and agents; ranked #1 open-weight on Artificial Analysis[18].

  10. Jan 9, 2026Hong Kong IPO

    Lists on the HKEX (00100); debut surges 70%+ to a peak market cap near $11.5B, days after rival Zhipu[1][14].

  11. Feb 2026Anthropic accusation

    Anthropic accuses MiniMax (with DeepSeek, Moonshot) of distilling Claude via fraudulent accounts — an unproven claim[4].

🗓️
Dates and figures are compiled from MiniMax's IPO prospectus (as reported by Chinese and English press) and the company's own model announcements. Later model-version dates beyond M2 (Oct 2025) were excluded where they could not be confirmed at a primary source.
Market & Industry Structure

China's model race — and a global consumer market

MiniMax sits in two markets at once: the hyper-competitive Chinese foundation-model race, and the global consumer-AI app market where most of its money is actually made.

Foundation modelsConsumer AI apps

The Chinese model market is defined by a price war and rapid open-weight commoditization — led by DeepSeek and the cloud giants — that crushes API margins[10]. MiniMax's answer is to compete on efficiency and to monetize overseas consumers (73% of revenue) rather than domestic API alone[15].

Two overlapping arenas

1) The foundation-model race.China's “six tigers” (Zhipu, Moonshot, MiniMax, Baichuan, 01.AI, StepFun) all reached unicorn status by early 2024, with a combined valuation estimated above RMB 100B by mid-2024[8]. The category is capital-intensive and increasingly commoditized: models are released open-weight and undercut on price almost as soon as they ship.

2) The consumer-AI app market. This is where MiniMax differs. Its revenue comes mainly from consumer subscriptions and in-app purchases on Hailuo (video) and Talkie/Xingye (companions), with 73% earned outside China across 200+ countries[15]. That makes it as much a global B2C app company as a Chinese model lab — and exposes it to app-store and content-moderation regimes worldwide.

The structural force: a price war meets open weights

Since 2024 a Chinese LLM price war — and the open-sourcing of strong models — has driven the price of raw intelligence toward zero. MiniMax frames this as a reason its efficiency matters: it reports training-cloud spend falling from over 1300% of revenue in 2023 to 266.5% in 2025 as its architecture matured[10]. The same force, though, means its core models are hard to charge for directly — a tension at the heart of the strategy section.

Tailwinds

  • Explosive global demand for consumer AI (video, voice, companions) where MiniMax already leads overseas[15].
  • Efficiency gains compounding — cloud cost ratio collapsing as models mature[10].
  • Multimodal breadth (text, speech, video) positions it across several fast-growing app categories[9].

Headwinds

  • A domestic price war and open-weight releases push the price of base models toward zero[10].
  • Compute supply is constrained by US chip-export controls — a hard cap on scale[31].
  • Crowded field — six tigers plus DeepSeek and big-tech model arms chase the same users and benchmarks[8].
📐
Market structure here is drawn from Chinese and English reporting on the “six tigers” and from MiniMax's prospectus commentary; combined-valuation figures are press estimates, not audited totals.
Business Model & Economics

Consumer apps fund the models

MiniMax monetizes mainly through consumer subscriptions and in-app purchases on Hailuo and Xingye, with a growing enterprise API platform — an unusually B2C revenue mix for a foundation-model lab.

~71% consumer73% overseas

In 9M 2025, roughly two-thirds of revenue came from two consumer apps — Hailuo AI (32.6%) and Xingye (35.1%) — with the enterprise open platform at 28.9% (+160% YoY)[11]. Gross margin has turned positive and is climbing (−24.7% → 12.2% → 23%+) as training-cloud costs fall[12].

Where the revenue comes from

MiniMax runs a hybrid model. The C-end (consumer) business sells subscriptions and in-app purchases on AI-native apps — Hailuo AI for video creation and Xingye/Talkie for AI companions — and is the majority of revenue. The B-end is an open API platform selling model access to enterprises and developers, growing fast (+160% YoY) off a smaller base[11].

9M'25
  • Xingye / companions (35.1%)35%
  • Hailuo AI / video (32.6%)33%
  • Enterprise open platform (28.9%)29%
  • Other (~3.4%)3%

Revenue mix per MiniMax's prospectus as reported by Wall Street CN[11]. Consumer apps (Hailuo + Xingye) are ~68%; the enterprise platform is the fastest-growing line.

23%+
gross margin, 9M 2025
from −24.7% in 2023 [12]
73%
of revenue from overseas
200+ countries [15]
+160%
enterprise platform growth (YoY)
B-end open API [11]
~1%
of US-model cost (abab claim)
within ~5% on quality [13]

The economics: efficiency vs. commoditization

The bull read is that MiniMax has found an efficient way to turn cheap intelligence into consumer revenue: gross margin is rising, sales spend actually fell~40% in 2025 (suggesting organic pull), and abab models are marketed as within ~5% of leading US models “at around 1% of the cost”[12][13]. The bear read is that the underlying product — model intelligence — is being commoditized by open weights and a price war, so the consumer apps must carry the whole business while the API line races to the bottom[10][19].

Why the model works

  • A real B2C franchise: ~68% of revenue from consumer apps people actually pay for[11].
  • Margins improving fast as cloud/training costs fall relative to revenue[12][10].
  • Cost leadership — near-frontier quality at a fraction of the cost lets it price aggressively[13].
  • Overseas mix diversifies away from the brutal domestic API market[15].

Why the model is fragile

  • Still unprofitable; absolute revenue is small relative to the ~$11B valuation[30].
  • Open-weight releases and free API trials undercut direct model monetization[19].
  • Consumer-app revenue depends on app stores and on companion apps with safety scrutiny[29].
  • Heavy reliance on two apps (Hailuo + Xingye) for ~68% of revenue is concentration risk[11].
⚖️
The core tension:MiniMax gives its models away to win distribution, then must recapture the value through consumer apps. Whether that funnel out-earns the cost of staying at the frontier is the business model's central open question.
Competitive Landscape & Positioning

A 'six tiger' that went global

MiniMax competes with China's other AI tigers, DeepSeek, and big-tech model arms — but positions itself apart through overseas consumer reach and open-weight, low-cost models.

Zhipu · Moonshot · DeepSeekByteDance · Alibaba

MiniMax's distinctive position is the open-weight + overseas-consumer corner: rivals like Zhipu and Moonshot are strong but more domestic/enterprise-tilted, while DeepSeek set the open-weight, low-price pace[14][15]. The race is close — capability gaps between open models are narrow and shrinking[16].

Who MiniMax competes with

  • Zhipu AI (Z.ai) — Beijing; GLM models; listed on the HKEX days before MiniMax as the “world's first LLM stock”[14].
  • Moonshot AI (Kimi) — Beijing; raised $500M (Dec 2025) at ~$4.3B and stayed private; strong on agentic/coding[14][21].
  • DeepSeek — set the open-weight, ultra-low-price benchmark that reshaped the market.
  • Baichuan, 01.AI, StepFun — the other “tigers,” with varying enterprise/consumer tilts[8].
  • Big-tech model arms — Alibaba (Qwen), ByteDance (Doubao), Tencent — with distribution and compute MiniMax can't match.
  • Global frontier labs — OpenAI, Anthropic, Google — the quality bar MiniMax benchmarks against[16].

Positioning: openness vs. go-to-market

One way to read the field is openness (closed vs. open-weight) against go-to-market (enterprise/domestic vs. global consumer). MiniMax sits in the open-weight, global-consumer quadrant — shared most closely with DeepSeek on openness but distinct on its consumer-app franchise.

Chinese AI model players — openness × go-to-market
Closed / proprietaryOpen-weightEnterprise / domesticGlobal consumerMiniMaxDeepSeekMoonshot (Kimi)Zhipu (GLM)Alibaba (Qwen)ByteDance (Doubao)

MiniMax: Open-weight M-series + 73% overseas consumer revenue

Five forces on China's consumer-AI model market

China AI models + consumer apps
Competitive rivalryHigh pressure. Six tigers + DeepSeek + Alibaba/ByteDance/Tencent model arms all chase the same benchmarks and users[8][14].

Why MiniMax stands out

  • The clearest overseas consumer franchise among the tigers — 73% international revenue[15].
  • Top-ranked open models (M2 #1 open-weight) give it developer mindshare and credibility[18].
  • Multimodal breadth (video, voice, companions) spreads it across several app categories[9].

Why rivals pressure it

  • DeepSeek set a price/openness bar that erodes everyone's model monetization[10][19].
  • Big-tech rivals (Alibaba, ByteDance) have compute and distribution MiniMax lacks[8].
  • Zhipu's earlier IPO and Moonshot's funding keep the capital race intense[14][21].
🧭
Peer positions are interpretive (analysis, not disclosed market shares) and drawn from the cited reporting; the 2×2 is a qualitative map, not a precise ranking.
Strategy & Moats

Win on efficiency, distribute by going open

MiniMax's stated strategy is frontier capability at a fraction of the cost, via novel attention architectures — then open-weight releases to seed developer adoption and consumer apps to monetize.

Lightning/linear attentionOpen-weight M-series

The technical bet is efficiency: a hybrid lightning/linear-attention + MoE architecture that makes long-context reasoning cheap. M1 trained its RL stage for just $534,700[17], and M2 (10B active params) ranked #1 open-weight on Artificial Analysis[18]. The catch: open weights are a distribution moat, not a pricing one.

Stated vs. revealed strategy

Stated: build the most capable models per dollar and make them broadly available. Revealed: use cheap, open models to win developer mindshare and feed a portfolio of global consumer apps that actually collect revenue. M1 was the first open-weight large-scale hybrid-attention reasoning model with a 1M-token context; M2 (MIT-licensed) targets coding and agentic workflows with only 10B active of 230B parameters for fast, cheap inference[17][18].

512 H800s for three weeks, with a rental cost of just $534,700.
MiniMax · on the M1 reinforcement-learning stage · Jun 2025 · source

The capability case, in benchmarks

Independent and reported benchmarks place MiniMax near the open-weight frontier — strong on agentic tool use and coding, the capabilities most useful for products.

MiniMax model benchmarks (reported, higher = better)
M2 · SWE-bench Verified
69.4
M1 · SWE-bench
56.0
M2 · Terminal-Bench
46.3
M2 · BrowseComp
44.0

Scores from MiniMax's M1 announcement and the M2 GitHub card[23][18]. Benchmarks are partly self-reported; treat exact values as indicative, and note M2 ranked #1 among open-weight models on Artificial Analysis's composite index.

Sources of advantage — and how they could erode

The scale signals are real: Speech-02 has generated 2.2B+ hours of audio, and the M2-series became the first Chinese model to exceed 50B daily tokens by Feb 2026[20]. But the deepest tension in the whole company lives here — its best models are given away. Open weights win adoption and goodwill; they do not, by themselves, produce gross profit, and M2's API launched free[19].

Why the moat is real

  • A genuine efficiency edge (linear attention + MoE) that lowers cost per token of intelligence[17][10].
  • Developer mindshare from #1-ranked open models plus surging token usage[18][20].
  • A consumer-app funnel (Hailuo, Xingye) that converts model strength into paid usage[11].

Why it could erode

  • Open weights are copyable; today's lead is next quarter's baseline for rivals[19].
  • Free APIs and a price war make the core model hard to monetize directly[19][10].
  • Efficiency advantages can be matched; DeepSeek already pushed the field toward cheap, open models[10].
⚖️
The strategic bet: that being the cheapest credible frontier model — and the most widely adopted open one — funnels enough developers and consumers into paid apps to out-earn the cost of staying at the frontier. It is a coherent bet, and an unproven one.
Peer Comparison & Benchmarking

MiniMax vs. the other tigers

Among China's 'six tigers,' MiniMax is the overseas-consumer specialist and one of the first to list. Peers vary widely in funding, IPO status and go-to-market.

Estimates · mixed disclosure

Zhipu and MiniMax won the 2026 IPO race (Zhipu listed first, Jan 8; MiniMax Jan 9), while Moonshot raised $500M at ~$4.3B and stayed private[14][21]. MiniMax stands out for its 73% overseasconsumer revenue and top-ranked open models[15][18].

The comparables table

CompanyHQFlagshipStatusGo-to-market tiltEdge
MiniMaxShanghaiHailuo, M-series[5]HK IPO Jan 9 2026[1]Global consumer (73% intl)[15]Efficiency + open + overseas
Zhipu (Z.ai)BeijingGLM[14]HK IPO Jan 8 2026[14]Enterprise / government“First LLM stock”
Moonshot (Kimi)BeijingKimi[14]Private; $500M raise[21]Domestic consumer + agenticLong-context / coding
DeepSeekHangzhouV/R-seriesPrivateDeveloper / open-weightPrice + open-weight pace
Baichuan / 01.AI / StepFunBeijing / ShanghaiBaichuan / Yi / Step[8]Private[8]MixedVaries

Compiled from Chinese and English reporting on the six tigers[8][14][21]. Several peers disclose little; treat funding/status as as-of the cited dates.

MiniMax's own scorecard

  • Revenue: ~$79M FY2025 (+159%), small but fast-growing off a tiny base[24].
  • Overseas mix: 73% — the highest international tilt among the tigers[15].
  • Users: 236M cumulative; 27.6M MAU (Sep 2025) from 3.1M in 2023[27].
  • Models: M2 ranked #1 open-weight on Artificial Analysis; M1 first open hybrid-attention reasoner[18][17].
  • Pre-IPO: ~$1.5B raised over 7 rounds at ~$4B; Alibaba ~13.7%, miHoYo ~6.4%[22].
📊
Read this as directional positioning, not a league table — the tigers differ so much in disclosure, go-to-market and model focus that head-to-head revenue/valuation comparisons are only loosely apples-to-apples.
Financials & Growth

Fast growth, real losses, a confusing headline number

MiniMax's IPO prospectus is a rare audited financial window among the tigers, most of which stay private. It shows triple-digit growth, improving margins — and a giant GAAP loss that mostly isn't operating cash.

HKEX prospectusNewly public

Revenue grew to $53.4M in 9M2025 (+175%) and gross margin turned positive (23%+)[24][12]. The eye-popping $512M net loss is mostly non-cash: ~$1.4B of fair-value charges on preferred shares for the year; the adjusted loss was ~$186M, roughly flat versus 2024[25].

Revenue trajectory

From a 2023 base of just $3.46M, revenue jumped to $30.5M in 2024 (+782%) and roughly $79M in FY2025 (+159%), with 73% earned overseas[24][2].

MiniMax revenue (US$M, disclosed)
202320249M'25FY'25

Revenue per the prospectus as reported by 36Kr[24]. Figures are in US dollars as reported.

$512M
GAAP net loss, 9M 2025
~$1.4B FY fair-value charge [25]
~$186M
adjusted net loss, 9M 2025
≈ flat vs 2024 [25]
$180M
R&D spend, 9M 2025
the core cost line [26]
$1.05B
cash reserves (Sep 2025)
runway before the IPO raise [26]

Reading the loss honestly

The headline GAAP loss is inflated by accounting that is standard for a pre-IPO company: rising fair-value charges on convertible/preferred shares as the equity value climbed — a non-cash item that disappears after listing. The operating reality is better captured by the adjusted net loss of ~$186M on $53.4M of revenue, and by R&D of $180M as the dominant cost[25][26]. Even so, this is a business spending well ahead of revenue — it had ~$1.05B cash before the IPO topped it up[26].

Bull case on the numbers

  • Triple-digit revenue growth with margins flipping positive and rising[24][12].
  • The scary GAAP loss is mostly non-cash; the real burn (~$186M adj.) is more modest[25].
  • ~$1.05B cash pre-IPO plus fresh listing proceeds give multi-year runway[26].

Bear case on the numbers

  • Absolute revenue (~$79M) is tiny against an ~$11B market cap — a steep multiple[1][24].
  • Still losing money every quarter in a price war that caps pricing[30][10].
  • R&D must keep rising to stay at the frontier, pressuring the path to profit[26].
⚠️
Estimate & interpretation warning:figures are from MiniMax's IPO prospectus as reported by the press, not an independent audit reviewed here. The GAAP-vs-adjusted gap is large and easy to misread; both are shown so you can judge. Currency is US dollars as reported.
Risks & Challenges

Losses, lawsuits, and geopolitics

MiniMax's risks are unusually broad for its size: financial (cash burn), legal (copyright and distillation), product (companion-app safety), and structural (export controls).

LegalSafetyGeopolitics

The defining risks cluster around money, law and geopolitics: an adjusted loss of ~$186M in a price war[30]; a Hollywood copyright suit over Hailuo[28]; an Anthropic distillation accusation[4]; companion-app safety scrutiny[29]; and US export controls on compute[31]. None has yet broken growth — but each is live.

1. Copyright — the Hailuo lawsuit

On September 16, 2025, Disney, Universal and Warner Bros. Discovery sued MiniMax in U.S. federal court, alleging Hailuo AI “pirates and plunders”their copyrighted works “on a massive scale” — citing, for example, on-demand generation of Darth Vader. They seek an injunction and up to $150,000 per infringed work[28]. The case is part of a wider wave of studio litigation against generative-AI video tools and is unresolved.

[MiniMax] pirates and plunders Plaintiffs' copyrighted works on a massive scale.
Disney, Universal, Warner Bros. Discovery · complaint, C.D. Cal. · Sep 16, 2025 · source

2. Distillation — the Anthropic accusation

In February 2026, Anthropic accused MiniMax (alongside DeepSeek and Moonshot) of “distillation” — using fraudulent accounts to generate ~13MClaude exchanges targeting agentic coding and tool use, and said it observed MiniMax “redirect nearly half its traffic” to siphon a new Claude model[4]. The allegation is unproven and MiniMax did not publicly respond; Anthropic tied it to the case for tighter US chip-export controls[31].

3. Product safety — companion apps

MiniMax's companion apps have repeatedly drawn safety and content-moderation scrutiny: the Chinese app Xingye/星野 was pulled over explicit content, and Talkie was removed from the U.S. App Store in December 2024[29]. For a business that earns the majority of revenue from consumer apps, app-store and child-safety regimes are a direct operational risk.

4. Financial & structural

  • Cash burn — Chinese coverage candidly framed the pre-IPO company as needing the listing to refuel, despite 170%+ growth[30].
  • Price war — domestic API pricing and open-weight releases cap monetization[10].
  • Compute — US export controls constrain access to leading-edge chips, a hard limit on scaling[31].
  • Concentration — ~68% of revenue from two apps; loss of either would hurt[11].

SWOT

Strengths

  • Overseas consumer franchise — 73% intl revenue, 236M users[15][27]
  • Top-ranked low-cost open models (M2 #1 open-weight)[18]
  • Efficiency architecture lowering cost per token[17][10]

Weaknesses

  • Still loss-making; small revenue vs. valuation[30][24]
  • Core models given away — weak direct monetization[19]
  • Revenue concentrated in two consumer apps[11]

Opportunities

  • Global consumer AI (video, voice, companions, agents) still early[9]
  • Enterprise platform growing +160% YoY[11]
  • Developer adoption from open weights → paid usage[20]

Threats

  • Copyright + distillation litigation[28][4]
  • US chip-export controls on compute[31]
  • Price war + big-tech rivals with more scale[10][8]
🧭
Critical claims are attributed to specific filings and reporting (the studios' complaint, Anthropic via TechCrunch, Chinese financial press). The distillation accusation is unproven and MiniMax has not publicly answered it; it is presented as an allegation, not a finding.
Forward View

Three questions that decide MiniMax's next chapter

Not a prediction — a map of the decisive uncertainties. Where the evidence leans, we say so, and show the strongest counter.

ScenariosAs of June 7, 2026

MiniMax's future turns on three things: whether its efficiency + open model edge stays ahead of commoditization; whether its overseas consumer franchise converts to profit before the price war bites; and whether legal and geopolitical overhangs stay manageable[19][15][31].

Question 1 — Does the efficiency/open edge endure?

The bull case is that lightning-attention efficiency and #1-ranked open models keep MiniMax at the frontier per dollar, compounding developer adoption (50B+ daily tokens) into a platform[18][20]. The risk: open weights are copyable and the whole field is racing the same benchmarks, so a lead is temporary by construction[19]. Where the evidence leans: a real lead today, but a treadmill, not a wall.

Question 2 — Does overseas consumer convert to profit?

MiniMax is further along than most rivals on monetization — 73% overseas, gross margin rising, adjusted loss roughly flat while revenue tripled[15][12][25]. The bear scenario is that consumer apps with safety scrutiny and app-store dependence can't scale margin fast enough against a price war and rising R&D[29][10].

Question 3 — Do the overhangs stay manageable?

A studio copyright suit, an unproven distillation accusation, and US export controls each carry tail risk — to Hailuo's legality in Western markets, to the company's reputation, and to its compute supply[28][4][31]. So far none has dented growth, but any could reprice the stock.

Scenarios

The efficient global platform (bull)

  • Open-model adoption + overseas consumer apps compound into the first profitable independent Chinese AI lab[18][15].
  • Margins keep climbing as efficiency improves and R&D leverages across products[12][10].
  • Multimodal breadth (video, voice, agents) opens several large consumer markets[9].

The commoditized cash-burner (bear)

  • Open weights + price war keep models near-free, so apps must carry everything[19][10].
  • Litigation or export controls constrain Western reach or compute[28][31].
  • Bigger-funded rivals (Alibaba, ByteDance, Moonshot) out-spend or out-distribute it[8][21].
⚖️
The one-line synthesis:MiniMax has done something unusual for a Chinese AI lab — built an overseas consumer-AI franchise (73% of revenue) and top-ranked open models on a fraction of rivals' spend — but it is priced as if that already adds up to durable profit it has not yet shown. The next two years are about whether efficiency and overseas reach compound faster than commoditization, litigation and geopolitics erode them. Reasonable people genuinely disagree, and this case study is built to let you weigh it yourself.
Methodology & Limitations

How this was built — and where it may be wrong

An honest account of sourcing, native-language research, frameworks, what is disclosed vs. estimated, and the as-of date after which this goes stale.

33 sources55% ChineseNeutral by design

How the research was done

This study was assembled by fan-out web research in both English and Chinese: searching, fetching and reading primary and secondary sources, then transcribing each load-bearing claim into a citation manifest with a tier, confidence and stance. Because MiniMax is a Chinese company whose home market and richest coverage are Chinese-language, a substantial share of sources are Chinese (18 of 33, ~55%), including disconfirming coverage; original-language text is shown alongside translations. Sources span MiniMax's own model pages and GitHub (Tier 1), reputable Chinese and English press (财联社, 华尔街见闻, 36氪, 澎湃, 新浪财经, 21世纪经济报道, TechNode, Variety, TechCrunch; Tier 2), and tertiary references (Wikipedia; Tier 3).

Frameworks used

  • Pyramid Principle — answer-first executive summary framing the open questions and balance of evidence.
  • Five Forces — structure of China's consumer-AI model market.
  • 2×2 positioning — openness × go-to-market across peers.
  • Peer benchmarking — MiniMax vs. Zhipu, Moonshot, DeepSeek and others.
  • Unit economics — revenue mix, gross-margin trajectory, efficiency.
  • SWOT — applied even-handedly in Risks.

Disclosed vs. estimated

MiniMax is newly public, so its IPO prospectus is the primary financial source — but the figures here are as reported by the press, not an independent audit reviewed in this study. Benchmark scores are partly self-reported by MiniMax. Valuations, peer figures and market-size context are estimates as of varying dates. The GAAP vs. adjusted loss distinction is large and important.

⚠️
Where this case study may be wrong:
  • Financials are from the prospectus as reported by media; exact figures (esp. the GAAP vs. ~$186M adjusted loss) could be misstated in secondary coverage.
  • Benchmark numbers (SWE-bench, Terminal-Bench, “#1 open-weight”) are partly self-reported and shift as rivals release new models.
  • The Anthropic distillation accusation is unproven; MiniMax has not publicly responded. It is an allegation, not a finding.
  • The copyright suit is ongoing; no liability has been established.
  • Valuation/market-cap figures are point-in-time and volatile for a freshly listed stock; pre-IPO round details are press estimates.
  • Model-version dates beyond M2 (Oct 2025) were excluded where unconfirmed; this is a snapshot as of June 7, 2026.

Neutrality commitment

This is a compilation meant to let you reach your own conclusion, not an argument for or against MiniMax. Each section presents supporting and countervailing evidence; positive and negative claims are held to the same sourcing standard. The achieved stance mix across sources is 14 supporting · 9 critical · 10 neutral (5 Tier-1, 23 Tier-2, 5 Tier-3).

🧭
Independent and not affiliated with, sponsored by, or endorsed by MiniMax (稀宇科技), and not investment advice — no rating, price target, or recommendation to buy or sell any security. As of June 7, 2026.
Sources

Full bibliography

Every load-bearing claim on this site links here. Each source was fetched during research; grouped by section, with tier, stance, language and confidence shown.

33 sources5 Tier-123 Tier-25 Tier-3
📊
Stance mix: 14 supporting · 9 critical · 10 neutral. Language: 18 Chinese-language sources (~55%), with original text shown. Tiers:Tier-1 = primary (MiniMax model pages & GitHub); Tier-2 = reputable secondary (财联社, 华尔街见闻, 36氪, 澎湃新闻, 新浪财经, 21世纪经济报道, TechNode, Variety, TechCrunch); Tier-3 = tertiary (Wikipedia), used for context or widely-reported facts and labeled as estimates where relevant. MiniMax is newly public; financial figures are from its IPO prospectus as reported by the press.

Executive Summary

  1. MiniMax listed on the HKEX (00100) on Jan 9, 2026 at HK$165, surged 70%+ on debut, and briefly topped a ~HK$90B (~$11.5B) market cap, raising ~HK$5.54B (~$710M).

    Offer price: HK$165 ... Debut gain: Surged more than 70% ... Market cap at peak: Above HK$90 billion ($11.5 billion).

    https://technode.com/2026/01/09/mihoyo-backed-ai-firm-minimax-jumps-on-hong-kong-debut-market-value-tops-11-5-billion/
  2. [2]华尔街见闻 — MiniMax首次披露业绩Tier 2neutralHigh confidenceZH

    For 9M ended Sep 30, 2025 MiniMax reported revenue of $53.4M (+~175% YoY) but a net loss of $512M; >70% of revenue is from overseas.

    2025 (9 months): $53.437 million (174.7% YoY) ... Adjusted net losses: $186M (2025 Q1-Q3) ... International revenue: 73.1% (2025 Q1-Q3).original · zh:2025年前九个月营收5343.7万美元,同比增长174.7%;经调整净亏损1.86亿美元;海外收入占比73.1%。

    https://wallstreetcn.com/articles/3761813
  3. MiniMax is one of China's AI 'six tigers' (六小虎) and went from founding (Dec 2021) to IPO in ~4 years, among the fastest globally; founder Yan Junjie is a former SenseTime VP.

    From founding to IPO the shortest-ever path among AI companies; 2.12 billion individual users across 200+ countries as of Sep 30, 2025.original · zh:成立于2022年初……从成立到IPO历时最短的AI公司;截至2025年9月30日,逾2.12亿名个人用户。

    https://m.cls.cn/detail/2235733
  4. Anthropic accused MiniMax (with DeepSeek and Moonshot) in Feb 2026 of 'distillation' — generating ~13M Claude exchanges via fraudulent accounts to siphon agentic/coding capability — an unproven allegation MiniMax did not publicly answer.

    13 million exchanges targeted agentic coding and tool use ... MiniMax ... redirected nearly half its traffic to siphon capabilities from the latest Claude model.

    https://techcrunch.com/2026/02/23/anthropic-accuses-chinese-ai-labs-of-mining-claude-as-us-debates-ai-chip-exports/

Company & Timeline

  1. [5]Wikipedia — MiniMax GroupTier 3neutralHigh confidence

    MiniMax was founded Dec 2021 in Shanghai by Yan Junjie (CEO) with co-founders from SenseTime; early products were Glow (Oct 2022, pulled in China 2023), Talkie (Jun 2023) and Xingye/星野 (Sep 2023).

    Founded: December 2021 ... Glow (October 2022) ... Talkie (June 2023) ... Xingye/星野 (September 2023).

    https://en.wikipedia.org/wiki/MiniMax_Group
  2. [6]Wikipedia — MiniMax GroupTier 3neutralHigh confidence

    Model cadence: abab 6.5 MoE (Apr 2024), MiniMax-01 (Jan 2025), Speech-02 (Apr 2025), open-weight M1 (Jun 2025), M2 (Oct 2025); the March 2024 round was a $600M raise led by Alibaba at a $2.5B valuation.

    March 2024: $600 million Series led by Alibaba at $2.5 billion valuation ... ABAB 6.5 (April 2024) ... MiniMax-01 (January 2025) ... M1 (June 2025).

    https://en.wikipedia.org/wiki/MiniMax_Group
  3. Founder Yan Junjie (闫俊杰), ~36, holds a PhD from the Chinese Academy of Sciences and was a SenseTime VP; the team is young (avg ~29) and R&D-heavy (~74% R&D).

    CEO Yan Junjie, age 36, PhD from Chinese Academy of Sciences, former VP at SenseTime; ~200 papers, 30,000+ citations.original · zh:闫俊杰,36岁,中科院博士,曾任商汤科技副总裁;发表论文约200篇,被引超3万次。

    https://m.thepaper.cn/newsDetail_forward_32366044

Market & Industry Structure

  1. [8]Wikipedia — Six AI tigersTier 3neutralMedium confidence

    China's 'six tigers' (Zhipu, Moonshot, MiniMax, Baichuan, 01.AI, StepFun) all reached unicorn status by early 2024, with a combined valuation estimated over RMB 100B by mid-2024; Alibaba and Tencent back several with capital and compute.

    by mid-2024, the combined valuation of the companies was estimated to exceed RMB 100 billion ... Alibaba Cloud and Tencent Cloud, providing both capital and compute resources.

    https://en.wikipedia.org/wiki/Six_AI_tigers
  2. [9]澎湃新闻 — 关于MiniMax你需要知道的几件事Tier 2supportingMedium confidenceZH

    MiniMax pursues a multimodal 'model-as-product' strategy across text, speech and video, and positions itself as among the global first tier on multimodal capability.

    Ranks fourth among pure-play large-model companies globally and is one of four companies with first-tier multimodal models globally.original · zh:在全球纯大模型公司中排名第四,是全球四家拥有第一梯队多模态模型的公司之一。

    https://m.thepaper.cn/newsDetail_forward_32366044
  3. The Chinese model market is intensely price-competitive (a 2024-25 API price war led by DeepSeek and cloud giants), pressuring everyone's economics; MiniMax leans on efficiency to compete.

    Training cloud expenses fell from over 1300% of revenue in 2023 to 266.5% in 2025.original · zh:训练云支出占收入的比例从2023年的逾1300%降至2025年的266.5%。

    https://m.thepaper.cn/newsDetail_forward_32366044

Business Model & Economics

  1. [11]华尔街见闻 — MiniMax首次披露业绩Tier 2neutralHigh confidenceZH

    MiniMax's revenue is mostly consumer (C-end) subscriptions/IAP from AI-native apps, plus a B-end open API platform; in 9M2025 Hailuo AI contributed 32.6% and Xingye 35.1% of revenue, with the enterprise platform 28.9% (+160% YoY).

    Hailuo AI and Xingye contributed 32.6% and 35.1% of Q1-Q3 2025 revenue respectively; enterprise platform contributed 28.9% (160% YoY growth).original · zh:海螺AI和星野分别贡献2025年前三季度收入的32.6%和35.1%;企业开放平台贡献28.9%,同比增长160%。

    https://wallstreetcn.com/articles/3761813
  2. Gross margin turned positive and is rising — from -24.7% (2023) to 12.2% (2024) to 23.3% (9M2025) / ~25.4% (FY2025) — as training-cloud costs fall relative to revenue.

    Gross margin: 25.4% (2024 was 12.2%; 2023 was negative).original · zh:毛利率25.4%(2024年为12.2%;2023年为负值)。

    https://36kr.com/p/3706674411549059
  3. [13]TechNode — MiniMax Hong Kong debutTier 2supportingMedium confidence

    abab models are positioned as near-frontier at a fraction of the cost — within ~5% of leading US models at roughly 1% of the cost, per company marketing cited at IPO.

    abab 6.5 model series delivers performance within 5% of leading US models 'at around 1% of the cost'.

    https://technode.com/2026/01/09/mihoyo-backed-ai-firm-minimax-jumps-on-hong-kong-debut-market-value-tops-11-5-billion/

Competitive Landscape & Positioning

  1. [14]Wikipedia — Six AI tigersTier 3neutralHigh confidence

    MiniMax competes head-to-head with the other 'six tigers' — notably Zhipu (HK-listed Jan 8, 2026 as the 'world's first LLM stock') and Moonshot/Kimi (stayed private) — plus DeepSeek and big-tech model arms.

    Zhipu AI ... 'World's First LLM Firm to Go Public' (January 2026) ... Moonshot AI ... 'Rules Out Quick IPO After Raising $500 Million'.

    https://en.wikipedia.org/wiki/Six_AI_tigers
  2. MiniMax differentiates on overseas consumer reach (73% of revenue international, 200+ countries) where Chinese rivals are more domestic-focused.

    Over 70% of revenue from overseas markets, products cover 200+ countries and regions.original · zh:超过70%的收入来自海外市场,产品覆盖200多个国家和地区。

    https://finance.sina.com.cn/roll/2025-12-23/doc-inhctsyv5655826.shtml
  3. [16]GitHub — MiniMax-M2Tier 1supportingMedium confidence

    Open-weight competition is fierce: M2 is benchmarked against DeepSeek, GLM (Zhipu) and Kimi (Moonshot) and frontier closed models like GPT-5 and Claude — a field where capability gaps are narrow and falling.

    Its composite score ranks #1 among open-source models globally ... SWE-bench Verified: 69.4 ... (Claude Sonnet 4.5: 63 AA Intelligence).

    https://github.com/MiniMax-AI/MiniMax-M2

Strategy & Moats

  1. [17]MiniMax — MiniMax-M1 announcementTier 1supportingHigh confidence

    MiniMax's technical bet is efficiency: a hybrid 'lightning/linear attention' + MoE architecture. M1 (Jun 16, 2025) is the first open-weight large-scale hybrid-attention reasoning model — 1M-token context, with an RL stage costing just $534,700.

    1 million token input window ... RL phase: 512 H800s for three weeks, with a rental cost of just $534,700.

    https://www.minimax.io/news/minimaxm1
  2. [18]GitHub — MiniMax-M2Tier 1supportingHigh confidence

    M2 (Oct 2025, MIT license) is a 230B-parameter MoE with only 10B active per token, built for coding and agentic workflows; independent evals ranked it #1 among open-weight models on the Artificial Analysis Intelligence Index.

    MoE model (230 billion total parameters with 10 billion active parameters) ... MIT license ... ranks #1 among open-source models globally.

    https://github.com/MiniMax-AI/MiniMax-M2
  3. [19]GitHub — MiniMax-M2Tier 1criticalMedium confidence

    Open-weight releases are a distribution moat but not a revenue one: M2's API was offered free for a limited time, reflecting how giving models away pressures direct monetization.

    The MiniMax-M2 API is 'free for a limited time' on the MiniMax Open Platform.

    https://github.com/MiniMax-AI/MiniMax-M2
  4. [20]36氪 — MiniMax首份财报Tier 2supportingMedium confidenceZH

    Scale signals for the model business: Speech-02 has generated 2.2B+ hours of audio, and M2-series text models surpassed 50B daily tokens (a first for a Chinese model) by Feb 2026.

    M2-series daily token consumption grew over 6x vs Dec 2025, the first Chinese model to exceed 50 billion daily tokens.original · zh:M2系列文本模型日Token消耗较2025年12月增长超6倍,成为首个日Token消耗突破500亿的中国模型。

    https://36kr.com/p/3706674411549059

Peer Comparison & Benchmarking

  1. Among peers, Zhipu listed in HK days before MiniMax; Moonshot AI raised $500M (Series C, Dec 2025) at a ~$4.3B valuation and stayed private — framing MiniMax's IPO as part of a 2026 'first LLM stock' race.

    Zhipu became the first large-model startup to formally launch IPO procedures (April 2024); MiniMax preliminarily preparing a Hong Kong listing (June 2025).original · zh:智谱率先于2024年4月正式启动IPO程序;MiniMax于2025年6月初步筹备港股上市。

    https://www.21jingji.com/article/20250715/herald/0c78cb8e4aefa9d2e44accb8d41e9f15.html
  2. Pre-IPO, MiniMax had raised ~$1.5B across 7 rounds at a ~$4B valuation (mid-2025), with Alibaba holding ~13.66% and miHoYo ~6.4%.

    Seven rounds raised over $1.5B; ~$4B valuation (July 2025); Alibaba holds ~13.66%.original · zh:经过7轮融资,累计吸金超15亿美元,估值约40亿美元(2025年7月);阿里巴巴持股约13.66%。

    https://finance.sina.com.cn/roll/2025-12-23/doc-inhctsyv5655826.shtml
  3. [23]MiniMax — MiniMax-M1 announcementTier 1supportingMedium confidence

    M1's reported benchmarks place it near the open-weight frontier: SWE-bench ~56% (just below DeepSeek-R1-0528's 57.6%) and TAU-bench agentic tool use leading all open-weight models, ahead of Gemini-2.5 Pro.

    TAU-bench (agent tool use): Leads all open-weight models, outperforms Gemini-2.5 Pro.

    https://www.minimax.io/news/minimaxm1

Financials & Growth

  1. [24]36氪 — MiniMax首份财报Tier 2supportingHigh confidenceZH

    Revenue scaled fast off a tiny base: $3.46M (2023) → $30.5M (2024, +782%) → $53.4M (9M2025, +175%) → ~$79M (FY2025, +159%).

    2023: $3.46M; 2024: $30.52M (+782%); FY2025: $79.04M (+158.9%).original · zh:2023年346万美元;2024年3052万美元(同比增长782%);2025年全年7903.8万美元(同比增长158.9%)。

    https://36kr.com/p/3706674411549059
  2. [25]36氪 — MiniMax首份财报Tier 2neutralHigh confidenceZH

    Losses are large but the GAAP figure overstates the operating burn: 9M2025 GAAP net loss was $512M, heavily inflated by ~$1.42B of non-cash fair-value losses on preferred shares; the adjusted net loss was ~$186M (9M2025), roughly flat vs 2024's $244M.

    Net loss reached $1.872B (partly due to $1.422B non-operating fair value losses). Adjusted net loss was $250M.original · zh:年内亏损18.72亿美元(其中含14.22亿美元非经营性公允价值损失);经调整净亏损2.5亿美元。

    https://36kr.com/p/3706674411549059
  3. [26]华尔街见闻 — MiniMax首次披露业绩Tier 2criticalHigh confidenceZH

    R&D was $180M in 9M2025 against just $53.4M revenue, and cash reserves were $1.046B at Sep 2025 — i.e. the business is spending well ahead of revenue and is not yet self-funding.

    R&D investment: $180M (2025 Q1-Q3); Cash reserves (Sept 2025): $1.046 billion.original · zh:研发投入2025年前三季度1.8亿美元;截至2025年9月现金储备10.46亿美元。

    https://wallstreetcn.com/articles/3761813
  4. [27]36氪 — MiniMax首份财报Tier 2supportingHigh confidenceZH

    User scale: 236M cumulative individual users and 214K enterprise customers/developers; MAU rose from 3.1M (2023) to 27.6M (Sep 2025).

    Cumulative personal users: 236M; enterprise customers and developers: 214K across 100+ countries.original · zh:累计个人用户2.36亿;企业客户和开发者21.4万家,覆盖100多个国家。

    https://36kr.com/p/3706674411549059

Risks & Challenges

  1. Disney, Universal and Warner Bros. Discovery sued MiniMax (Sep 16, 2025, C.D. Cal.) alleging Hailuo AI 'pirates and plunders' their works 'on a massive scale' (e.g., generating Darth Vader), seeking injunction and up to $150,000 per infringed work.

    pirates and plunders ... on a massive scale ... maximum statutory damages of $150,000 per infringed work, plus an injunction.

    https://variety.com/2025/digital/news/disney-warner-bros-discovery-nbcu-lawsuit-minimax-chinese-ai-company-1236520395/
  2. [29]Wikipedia — MiniMax GroupTier 3criticalMedium confidence

    MiniMax's companion apps (Talkie internationally, Xingye/星野 in China) have drawn safety and content-moderation criticism — Xingye was pulled in China over explicit content, and Talkie was removed from the U.S. App Store in Dec 2024.

    Xingye ... was pulled from app stores due to explicit content ... Talkie ... removed from the U.S. App Store in 2024.

    https://en.wikipedia.org/wiki/MiniMax_Group
  3. The business is deeply loss-making and cash-hungry, and Chinese coverage framed the pre-IPO company as a 'cash-burning beast' needing the listing to refuel — a candid domestic read on the economics.

    Adjusted net loss of $186M in 9M2025 despite 170%+ revenue growth; still unprofitable.original · zh:尽管收入同比增长超170%,2025年前三季度仍录得经调整净亏损1.86亿美元,尚未盈利。

    https://m.thepaper.cn/newsDetail_forward_32366044
  4. [31]TechCrunch — Anthropic accuses Chinese AI labsTier 2criticalMedium confidence

    Geopolitics is a structural risk: Anthropic argued the alleged distillation 'reinforces the rationale for export controls', tying MiniMax to the U.S.–China chip-export debate that constrains its compute supply.

    distillation attacks therefore reinforce the rationale for export controls.

    https://techcrunch.com/2026/02/23/anthropic-accuses-chinese-ai-labs-of-mining-claude-as-us-debates-ai-chip-exports/

Forward View

  1. [32]财联社 — MiniMax通过港交所聆讯Tier 2supportingMedium confidenceZH

    The bull case: a fast-listing, overseas-led consumer franchise (73% international, 27.6M MAU) plus top-ranked low-cost open models could compound into the first profitable independent Chinese AI lab.

    Among the fastest globally from founding to IPO; 2.12 billion cumulative individual users.original · zh:全球从成立到上市最快的AI公司之一;累计个人用户超2.12亿。

    https://m.cls.cn/detail/2235733
  2. The bear case: persistent losses, a domestic price war, open-weight commoditization of its core models, unresolved IP/distillation litigation and export-control exposure could outrun its narrowing path to profit.

    Overseas surging, but profit not yet arrived.original · zh:海外狂飙,盈利未至。

    https://36kr.com/p/3706674411549059