MiniMax FY2025 Earnings: $79M in Revenue, and a $1.9 Billion Loss That Isn't What It Looks Like

MiniMax Group Inc. — the Shanghai-based AI foundation model company behind Hailuo AI, Talkie, and the MiniMax-M2 model series — just reported its first full year of audited results as a public company: revenue nearly tripled to $79.0 million (+158.9% year over year), gross margin more than doubled to 25.4%, and yet the headline net loss under IFRS ballooned to $1.87 billion, up 302% from the prior year. Read only the bottom line and MiniMax looks like it is losing money faster than ever. Read the ledger and the opposite is true: $1.59 billion of that loss is a single non-cash line — the fair-value remeasurement of preferred shares that automatically converted to ordinary equity when the company listed on the Hong Kong Stock Exchange on January 9, 2026. Strip it out, and adjusted net loss grew just 2.7%, from $244.2 million to $250.9 million, while revenue grew 159%. That gap between the GAAP story and the operating story is the whole earnings release.

The Headline Numbers​

MiniMax reports under IFRS Accounting Standards, in USD, for the fiscal year ended December 31, 2025:

Two numbers in that table are doing all the work. Revenue growth of 158.9% against research and development growth of only 33.8% is the efficiency story — MiniMax is extracting far more commercial output per R&D dollar than a year ago. And selling and distribution expense actually fell 40.3% in dollar terms while revenue nearly tripled, which is the rarest combination in consumer AI: growth that is becoming less, not more, expensive to buy.

The fair value loss line is the distortion. It is a mechanical consequence of MiniMax's pre-IPO capital structure, not a reflection of the underlying business, and we unpack exactly how it works below.

Revenue Deep Dive: Two Businesses, One Model Stack​

MiniMax's $79.0 million in FY2025 revenue splits into two segments built on the same underlying foundation models:

AI-native products ($53.1M, +143%): This is MiniMax's consumer- and creator-facing product suite — Hailuo AI (video and multimodal generation, including the Hailuo 2.3 model with a faster "Fast" tier that cuts batch content-creation costs up to 50%), Talkie/Xingye (AI companion and social apps), MiniMax Audio (voice and music generation), and the newly monetizing MiniMax Agent. Talkie/Xingye alone was essentially the entire consumer business as recently as 2023; by FY2025 Hailuo AI had grown into the largest single product line, evidence that the multimodal push beyond text is translating into paying usage.

Open Platform and other AI-based enterprise services ($26.0M, +198%): This is the developer- and enterprise-facing API business — the same MiniMax-M2 model family sold as inference to third parties building their own products. The company reported more than 236 million cumulative users and 214,000 enterprise customers and developers across over 100 countries by year-end. More than 70% of total revenue was generated outside Chinese mainland, with the U.S., Singapore, and a long tail of other markets (221 distinct jurisdictions contributed revenue in FY2025, up from 134 a year earlier) doing the heavy lifting — a notable data point for a company headquartered in Shanghai.

The model quality behind both segments has been improving on a fast cadence: MiniMax-M2, released in Q4 2025, became the first Chinese model on OpenRouter to exceed 50 billion tokens of daily consumption and topped HuggingFace's global trending leaderboard; its February 2026 successor, M2.5, set a new record on the SWE-Bench Verified coding benchmark. Average daily token consumption of the M2 series has grown roughly 6x from December 2025 to February 2026 alone — the revenue numbers above likely understate where the business is now, several months after the fiscal year closed.

The Margin Story​

MiniMax's margin structure across its short public history shows an operating-leverage curve that is unusually steep, because the company started from zero:

In FY2023, MiniMax spent more than 20x its revenue on R&D alone — the textbook shape of a foundation-model company still building the product before it has customers to fund it. By FY2025, R&D is "only" 3.2x revenue, and selling and distribution costs have collapsed from 6.6x revenue to two-thirds of revenue, in a single two-year window. Gross margin improved 13.2 percentage points in FY2025 alone, driven by better model and system efficiency and smarter infrastructure allocation — cost of sales (mostly third-party cloud inference spend) grew 120% while the revenue it supported grew 159%. Every one of these ratios is still compressing at a fast clip; none of them is close to a mature software company's economics yet, but the trajectory is what a monetization inflection looks like in the numbers.

Administrative expense is the one line moving the "wrong" way, up 155.9% to $36.8 million, driven by increased headcount, higher share-based compensation for administrative staff, and $6.9 million of one-time listing expenses tied to the Hong Kong IPO that had no FY2024 equivalent. Strip out the listing cost and underlying G&A growth is closer to 100% — still elevated, but consistent with a company that just built out the finance, legal, and investor-relations functions a public listing requires.

The $1.9 Billion Question: Is MiniMax Really Losing More Money?​

No. This is the section that matters most for reading MiniMax's numbers correctly.

Before its IPO, MiniMax — like almost every venture-backed startup — raised capital through convertible redeemable preferred shares. Under IFRS (IAS 32), preferred shares with a redemption feature are not "equity" the way common stock is; they are classified as a financial liability, carried on the balance sheet at fair value, and remeasured every reporting period. Every time MiniMax's implied valuation rose in a funding round, the fair value of that liability rose with it — and that increase runs through the income statement as an expense, even though no cash changed hands and nothing about the operating business changed. Here is how that liability grew across MiniMax's four years as a private, then newly-public, company:

The FY2025 spike is enormous because MiniMax's implied valuation jumped sharply in the run-up to its Hong Kong listing (which ultimately priced the IPO around a $4 billion valuation and then saw shares more than double on their January 9, 2026 debut) — the better the company's prospects looked to public-market investors, the larger the non-cash loss IFRS required MiniMax to book on its own preferred shares. This is a mechanical artifact of pre-IPO accounting, not a signal of deteriorating unit economics, and it is also a one-time phenomenon: those preferred shares were automatically converted to ordinary shares and re-designated from a liability to equity at Listing, nine days into FY2026. There will be no FY2026 fair value loss line, because there will be no more preferred-share liability to remeasure.

MiniMax's own non-IFRS reconciliation makes the adjustment explicit:

A company whose GAAP loss grew 302% but whose adjusted loss grew 2.7% — against revenue growth of 159% — is a company getting dramatically more efficient, not less.

What Wall Street Thinks​

Coverage of MiniMax remains thin relative to U.S. megacaps, but the sell side that has initiated is constructive. Goldman Sachs analyst Ronald Keung maintains a Buy rating with a HK$860 price target, and Citi has also maintained a Buy rating, with a HK$533 price target. The stock's round trip has already been volatile — shares more than doubled on their Hong Kong debut in January 2026 before giving back much of that premium as the broader Chinese AI-listing wave (MiniMax was quickly followed onto HKEX by rivals including Zhipu AI) diluted investor attention across several similarly-positioned names at once. That volatility is itself a reasonable proxy for how differently GAAP and adjusted numbers can be read by a market still calibrating how to value a loss-making, pre-profit foundation-model company against a backdrop of intensifying China–U.S.–open-source model competition.

Tracking a Foundation Model Company in Plain Text​

The clearest way to see the fair-value distortion for what it is: put MiniMax's complete financial history into Beancount, the plain-text double-entry accounting system, where every dollar of the loss has to reconcile to a specific line item and there is nowhere for a non-cash charge to hide.

Here is the FY2025 income statement exactly as it is modeled in the ledger — Income accounts are negative (credit balances), Expenses are positive (debit balances), and the fair value loss gets its own dedicated account precisely because it is too large and too temporary to bury inside "other expenses":

; FY2025 Income Statement — fiscal year ended December 31, 2025
; 1 MUSD = USD 1,000,000  |  All figures in millions USD
; Check: −79.038 + (−39.634) + 58.959 + 252.771 + 88.709 + 1589.850 + 0 + (−1871.617) = 0 ✓
 
2025-12-31 * "MiniMax Group Inc." "FY2025 Income Statement"
  Income:Revenue                              -79.038 MUSD  ; revenue earned (credit)
  Income:OtherNet                             -39.634 MUSD  ; net other income (credit)
  Expenses:CostOfRevenue                        58.959 MUSD  ; cost incurred (debit)
  Expenses:ResearchAndDevelopment               252.771 MUSD ; cost incurred (debit)
  Expenses:SellingGeneralAdministrative          88.709 MUSD ; S&D + Admin combined (debit)
  Expenses:FairValueLossOnFinancialLiabilities 1589.850 MUSD ; non-cash preferred-share remeasurement (debit)
  Expenses:IncomeTax                              0 MUSD
  Equity:Adjustments                         -1871.617 MUSD  ; net loss offset (RE set by balance assertion)

And here is the balance sheet item that explains everything else — the preferred-share liability climbing across four fiscal year-end snapshots, right up until the moment it disappears:

; Convertible redeemable preferred shares — pad from Equity:Adjustments, balance on period end
2022-12-31 balance Liabilities:Current:ConvertibleRedeemablePreferredShares  -145.175 MUSD
2023-12-31 balance Liabilities:Current:ConvertibleRedeemablePreferredShares  -629.001 MUSD
2024-12-31 balance Liabilities:Current:ConvertibleRedeemablePreferredShares -1581.949 MUSD
2025-12-31 balance Liabilities:Current:ConvertibleRedeemablePreferredShares -3597.566 MUSD
; Converts to ordinary equity at Listing, 2026-01-09 — nine days into FY2026.
; The account effectively goes to zero the moment the company stops being private.

By the FY2025 balance-sheet date, that single liability is larger than MiniMax's total assets ($1,088.4 million), which is why the company's own filings describe its financial position as "net liabilities" of $2,648.2 million — a number that reads alarmingly until you know it evaporates by conversion rather than by repayment. Meanwhile, the assets side of the balance sheet tells a much healthier story: cash and short-term investments (cash $507.6 million + short-term investments $452.3 million) totaled roughly $960 million at year-end, funded by the same rounds of preferred-share capital that created the liability — MiniMax is sitting on a large, liquid war chest, not running out of runway.

The complete four-year ledger — FY2022 through FY2025, balance sheet and income statement, fully reconciled to the company's IPO prospectus and its first annual report as a listed company — is open and auditable:

The Four-Year Arc: From Zero Revenue to a Public Listing​

MiniMax's entire operating history as a commercial company fits inside four fiscal years:

MiniMax generated no revenue at all in 2022 — it was purely an R&D lab building foundation models. By 2025, it had 236 million cumulative users, 214,000 enterprise customers, and $79 million in annual revenue growing at triple-digit rates. Adjusted net loss grew far more slowly than either revenue or the balance sheet's cash pile, which is the shape a bet on foundation-model scaling is supposed to take: spend aggressively and raise capital aggressively while the technology and the market are both moving, and let revenue eventually catch up to the cost base. Whether it fully catches up — turning a $250.9 million annual adjusted loss into a profit — is the multi-year question the public markets are now pricing daily.

The Verdict: Bull vs. Bear​

The Bull Case:

• Revenue grew 158.9% in FY2025 on top of 782% growth in FY2024 — two consecutive years of triple-digit-plus expansion is a rare combination even among well-funded AI labs
• Selling and distribution expense fell 40.3% in absolute dollars while revenue nearly tripled, evidence that organic growth (user referrals, product virality in Talkie/Xingye and Hailuo AI) is displacing paid acquisition
• Adjusted net loss grew only 2.7% against 158.9% revenue growth — the single clearest efficiency signal in the entire release
• MiniMax-M2 became the first Chinese model to cross 50 billion daily tokens on OpenRouter and topped HuggingFace's trending leaderboard, with the February 2026 successor M2.5 setting new coding-benchmark records — evidence the model quality gap with U.S. labs is narrowing on a fast cadence
• Roughly $960 million of cash and short-term investments at year-end, funded by successful pre-IPO rounds and a January 2026 listing that doubled on debut, gives ample runway to keep scaling compute and headcount

The Bear Case:

• The IFRS net loss of $1.87 billion, while mostly a non-cash artifact, is still what shows up in GAAP-based screens and index inclusion criteria, and could weigh on institutional ownership until a full post-conversion fiscal year is reported
• R&D expense still runs at over 3x revenue; even after dramatic improvement, MiniMax needs several more years of this trajectory before the cost base looks anything like a mature software company's
• Administrative expenses grew 155.9%, and while listing costs explain part of it, rising headcount-driven G&A is a trend worth watching for a company still years from profitability
• The stock has already been volatile post-listing — more than doubling on debut and then giving back much of that gain — as investors sort MiniMax's fundamentals from a broader wave of similarly-timed Chinese AI IPOs (including Zhipu AI) competing for the same pool of capital
• Foundation-model economics are brutally competitive: MiniMax competes directly with DeepSeek, Zhipu AI, and every major U.S. lab on model quality and API pricing simultaneously, and a single competitor undercutting Open Platform pricing could compress the 25.4% gross margin MiniMax just fought hard to build

Our Take: The GAAP headline — a $1.87 billion annual loss — is the least useful number in this entire earnings release, and investors who stop there will misprice the company. The number that matters is the $250.9 million adjusted net loss growing just 2.7% while revenue grew 159%, alongside a selling and distribution line that is shrinking in absolute dollars. That combination describes a company converting an R&D-heavy, all-in AI infrastructure bet into a real, monetizing product business faster than its cost base is expanding — precisely the operating leverage story that made Microsoft's Intelligent Cloud segment or Circle's stablecoin reserve income worth modeling in the first place. The preferred-share overhang was always a pre-IPO artifact by design; it is now gone. What is left, starting with FY2026, is the first fiscal year where MiniMax's income statement will finally look like what the business actually is.